When Optimal is the Enemy of Good: High-Budget Differential Privacy for Medical AI

Stay Ahead, Stay ONMINE

When Optimal is the Enemy of Good: High-Budget Differential Privacy for Medical AI

Imagine you’re building your dream home. Just about everything is ready. All that’s left to do is pick out a front door. Since the neighborhood has a low crime rate, you decide you want a door with a standard lock — nothing too fancy, but probably enough to deter 99.9% of would-be burglars. Unfortunately, the local homeowners’ association (HOA) has a rule stating that all front doors in the neighborhood must be bank vault doors. Their reasoning? Bank vault doors are the only doors that have been mathematically proven to be absolutely secure. As far as they’re concerned, any front door below that standard may as well not be there at all. You’re left with three options, none of which seems particularly appealing: Concede defeat and have a bank vault door installed. Not only is this expensive and cumbersome, but you’ll be left with a front door that bogs you down every single time you want to open or close it. At least burglars won’t be a problem! Leave your house doorless. The HOA rule imposes requirements on any front door in the neighborhood, but it doesn’t technically forbid you from not installing a door at all. That would save you a lot of time and money. The downside, of course, is that it would allow anyone to come and go as they please. On top of that, the HOA could always close the loophole, taking you back to square one. Opt out entirely. Faced with such a stark dilemma (all-in on either security or practicality), you choose not to play the game at all, selling your nearly-complete house and looking for someplace else to live. This scenario is obviously completely unrealistic. In real life, everybody strives to strike an appropriate balance between security and practicality. This balance is informed by everyone’s own circumstances and risk analysis, but it universally lands somewhere between the two extremes of bank vault door and no door at all. But what if instead of your dream home, you imagined a medical AI model that has the power to help doctors improve patient outcomes? Highly-sensitive training data points from patients are your valuables. The privacy protection measures you take are the front door you choose to install. Healthcare providers and the scientific community are the HOA. Suddenly, the scenario is much closer to reality. In this article, we’ll explore why that is. After understanding the problem, we’ll consider a simple but empirically effective solution proposed in the paper Reconciling privacy and accuracy in AI for medical imaging [1]. The authors propose a balanced alternative to the three bad choices laid out above, much like the real-life approach of a typical front door. The State of Patient Privacy in Medical AI Over the past few years, artificial intelligence has become an ever more ubiquitous part of our day-to-day lives, proving its utility across a wide range of domains. The rising use of AI models has, however, raised questions and concerns about protecting the privacy of the data used to train them. You may remember the well-known case of ChatGPT, just months after its initial release, exposing proprietary code from Samsung [2]. Some of the privacy risks associated with AI models are obvious. For example, if the training data used for a model isn’t stored securely enough, bad actors could find ways to access it directly. Others are more insidious, such as the risk of reconstruction. As the name implies, in a reconstruction attack, a bad actor attempts to reconstruct a model’s training data without needing to gain direct access to the dataset. Medical records are one of the most sensitive kinds of personal information there are. Although specific regulation varies by jurisdiction, patient data is generally subject to stringent safeguards, with hefty fines for inadequate protection. Beyond the letter of the law, unintentionally exposing such data could irreparably damage our ability to use specialized AI to empower medical professionals. As Ziller, Mueller, Stieger, et al. point out [1], fully taking advantage of medical AI requires rich datasets comprising information from actual patients. This information must be obtained with the full consent of the patient. Ethically acquiring medical data for research was challenging enough as it was before the unique challenges posed by AI came into play. But if proprietary code being exposed caused Samsung to ban the use of ChatGPT [2], what would happen if attackers managed to reconstruct MRI scans and identify the patients they belonged to? Even isolated instances of negligent protection against data reconstruction could end up being a monumental setback for medical AI as a whole. Tying this back into our front door metaphor, the HOA statute calling for bank vault doors starts to make a little bit more sense. When the cost of a single break-in could be so catastrophic for the entire neighborhood, it’s only natural to want to go to any lengths to prevent them. Differential Privacy (DP) as a Theoretical Bank Vault Door Before we discuss what an appropriate balance between privacy and practicality might look like in the context of medical AI, we have to turn our attention to the inherent tradeoff between protecting an AI model’s training data and optimizing for quality of performance. This will set the stage for us to develop a basic understanding of Differential Privacy (DP), the theoretical gold standard of privacy protection. Although academic interest in training data privacy has increased significantly over the past four years, principles on which much of the conversation is based were pointed out by researchers well before the recent LLM boom, and even before OpenAI was founded in 2015. Though it doesn’t deal with reconstruction per se, the 2013 paper Hacking smart machines with smarter ones [3] demonstrates a generalizable attack methodology capable of accurately inferring statistical properties of machine learning classifiers, noting: “Although ML algorithms are known and publicly released, training sets may not be reasonably ascertainable and, indeed, may be guarded as trade secrets. While much research has been performed about the privacy of the elements of training sets, […] we focus our attention on ML classifiers and on the statistical information that can be unconsciously or maliciously revealed from them. We show that it is possible to infer unexpected but useful information from ML classifiers.” [3] Theoretical data reconstruction attacks were described even earlier, in a context not directly pertaining to machine learning. The landmark 2003 paper Revealing information while preserving privacy [4] demonstrates a polynomial-time reconstruction algorithm for statistical databases. (Such databases are intended to provide answers to questions about their data in aggregate while keeping individual data points anonymous.) The authors show that to mitigate the risk of reconstruction, a certain amount of noise needs to be introduced into the data. Needless to say, perturbing the original data in this way, while necessary for privacy, has implications for the quality of the responses to queries, i.e., the accuracy of the statistical database. In explaining the purpose of DP in the first chapter of their book The Algorithmic Foundations of Differential Privacy [5], Cynthia Dwork and Aaron Roth address this tradeoff between privacy and accuracy: “[T]he Fundamental Law of Information Recovery states that overly accurate answers to too many questions will destroy privacy in a spectacular way. The goal of algorithmic research on differential privacy is to postpone this inevitability as long as possible. Differential privacy addresses the paradox of learning nothing about an individual while learning useful information about a population.” [5] The notion of “learning nothing about an individual while learning useful information about a population” is captured by considering two datasets that differ by a single entry (one that includes the entry and one that doesn’t). An (ε, δ)-differentially private querying mechanism is one for which the probability of a certain output being returned when querying one dataset is at most a multiplicative factor of the probability when querying the other dataset. Denoting the mechanism by M, the set of possible outputs by S, and the datasets by x and y, we formalize this as [5]: Pr[M(x) ∈ S] ≤ exp(ε) ⋅ Pr[M(y) ∈ S] + δ Where ε is the privacy loss parameter and δ is the failure probability parameter. ε quantifies how much privacy is lost as a result of a query, while a positive δ allows for privacy to fail altogether for a query at a certain (usually very low) probability. Note that ε is an exponential parameter, meaning that even slightly increasing it can cause privacy to decay significantly. An important and useful property of DP is composition. Notice that the definition above only applies to cases where we run a single query. The composition property helps us generalize it to cover multiple queries based on the fact that privacy loss and failure probability accumulate predictably when we compose several queries, be they based on the same mechanism or different ones. This accumulation is easily proven to be (at most) linear [5]. What this means is that, rather than considering a privacy loss parameter for one query, we may view ε as a privacy budget that can be utilized across a number of queries. For example, when taken together, one query using a (1, 0)-DP mechanism and two queries using a (0.5, 0)-DP mechanism satisfy (2, 0)-DP. The value of DP comes from the theoretical privacy guarantees it promises. Setting ε = 1 and δ = 0, for example, we find that the probability of any given output occurring when querying dataset y is at most exp(1) = e ≈ 2.718 times greater than that same output occurring when querying dataset x. Why does this matter? Because the greater the discrepancy between the probabilities of certain outputs occurring, the easier it is to determine the contribution of the individual entry by which the two datasets differ, and the easier it is to ultimately reconstruct that individual entry. In practice, designing an (ε, δ)-differentially private randomized mechanism entails the addition of random noise drawn from a distribution dependent on ε and δ. The specifics are beyond the scope of this article. Shifting our focus back to machine learning, though, we find that the idea is the same: DP for ML hinges on introducing noise into the training data, which yields robust privacy guarantees in much the same way. Of course, this is where the tradeoff we mentioned comes into play. Adding noise to the training data comes at the cost of making learning more difficult. We could absolutely add enough noise to achieve ε = 0.01 and δ = 0, making the difference in output probabilities between x and y virtually nonexistent. This would be wonderful for privacy, but terrible for learning. A model trained on such a noisy dataset would perform very poorly on most tasks. There is no consensus on what constitutes a “good” ε value, or on universal methodologies or best practices for ε selection [6]. In many ways, ε embodies the privacy/accuracy tradeoff, and the “proper” value to aim for is highly context-dependent. ε = 1 is generally regarded as offering high privacy guarantees. Although privacy diminishes exponentially with respect to ε, values as high as ε = 32 are mentioned in literature and thought to provide moderately strong privacy guarantees [1]. The authors of Reconciling privacy and accuracy in AI for medical imaging [1] test the effects of DP on the accuracy of AI models on three real-world medical imaging datasets. They do so using various values of ε and comparing them to a non-private (non-DP) control. Table 1 provides a partial summary of their results for ε = 1 and ε = 8: Table 1: Comparison of AI model performance across the RadImageNet [7], HAM10000 [8], and MSD Liver [9] datasets with δ = 8⁻⁷⋅10 and privacy budgets of ε = 1, ε = 8, and without DP (non-private). A higher MCC/Dice score indicates higher accuracy. Although providing strong theoretical privacy guarantees in the face of a worst-case adversary, DP significantly degrades model accuracy. The negative impact on performance is especially noticeable in the latter two datasets, which are considered small datasets. Image by the author, based on image by A. Ziller, T.T. Mueller, S. Stieger, et al from Table 3 in Reconciling privacy and accuracy in AI for medical imaging [1] (use under CC-BY 4.0 license). Even approaching the higher end of the typical ε values attested in literature, DP is still as cumbersome as a bank vault door for medical imaging tasks. The noise introduced into the training data is catastrophic for AI model accuracy, especially when the datasets at hand are small. Note, for example, the huge drop-off in Dice score on the MSD Liver dataset, even with the relatively high ε value of 8. Ziller, Mueller, Stieger, et al. suggest that the accuracy drawbacks of DP with typical ε values may contribute to the lack of widespread adoption of DP in the field of Medical Ai [1]. Yes, wanting mathematically-provable privacy guarantees is definitely sensible, but at what cost? Leaving so much of the diagnostic power of AI models on the table in the name of privacy is not an easy choice to make. Revisiting our dream home scenario armed with an understanding of DP, we find that the options we (seem to) have map neatly onto the three we had for our front door. DP with typical values of ε is like installing a bank vault door: costly, but effective for privacy. As we’ll see, it’s also complete overkill in this case. Not using DP is like not installing a door at all: much easier, but risky. As mentioned above, though, DP has yet to be widely applied in medical AI [1]. Passing up opportunities to use AI is like giving up and selling the house: it saves us the headache of dealing with privacy concerns weighed against incentives to maximize accuracy, but a lot of potential is lost in the process. It looks like we’re at an impasse… unless we think outside the box. High-Budget DP: Privacy and Accuracy Aren’t an Either/Or In Reconciling privacy and accuracy in AI for medical imaging [1], Ziller, Mueller, Stieger, et al. offer the medical AI equivalent of a regular front door — an approach that manages to protect privacy while giving up very little in the way of model performance. Granted, this protection is not theoretically optimal — far from it. However, as the authors show through a series of experiments, it is good enough to counter almost any realistic threat of reconstruction. As the saying goes, “Perfect is the enemy of good.” In this case, it is the “optimal” — an insistence on arbitrarily low ε values — that locks us into the false dichotomy of total privacy versus total accuracy. Just as a bank vault door has its place in the real world, so does DP with ε ≤ 32. Still, the existence of the bank vault door doesn’t mean plain old front doors don’t also have a place in the world. The same goes for high-budget DP. The idea behind high-budget DP is straightforward: using privacy budgets (ε values) that are so high that they “are near-universally shunned as being meaningless” [1] — budgets ranging from ε = 10⁶ to as high as ε = 10¹⁵. In theory, these provide such weak privacy guarantees that it seems like common sense to dismiss them as no better than not using DP at all. In practice, though, this couldn’t be further from the truth. As we will see by looking at the results from the paper, high-budget DP shows significant promise in countering realistic threats. As Ziller, Mueller, Stieger, et al. put it [1]: “[E]ven a ‘pinch of privacy’ has drastic effects in practical scenarios.” First, though, we need to ask ourselves what we consider to be a “realistic” threat. Any discussion of the efficacy of high-budget DP is inextricably tied to the threat model under which we choose to evaluate it. In this context, a threat model is simply the set of assumptions we make about what a bad actor interested in obtaining our model’s training data is able to do. Table 2: Comparison of threat models. For all three, we also assume that the adversary has unbounded computational ability. Image by A. Ziller, T.T. Mueller, S. Stieger, et al from Table 1 in Reconciling privacy and accuracy in AI for medical imaging [1] (use under CC-BY 4.0 license). The paper’s findings hinge on a calibration of the assumptions to better suit real-world threats to patient privacy. The authors argue that the worst-case model, which is the one typically used for DP, is far too pessimistic. For example, it assumes that the adversary has full access to each original image while attempting to reconstruct it based on the AI model (see Table 2) [1]. This pessimism explains the discrepancy between the reported “drastic effects in practical scenarios” of high privacy budgets and the very weak theoretical privacy guarantees that they offer. We may liken it to incorrectly assessing the security threats a typical house faces, wrongly assuming they are likely to be as sophisticated and enduring as those faced by a bank. The authors therefore propose two alternative threat models, which they call the “relaxed” and “realistic” models. Under both of these, adversaries keep some core capabilities from the worst-case model: access to the AI model’s architecture and weights, the ability to manipulate its hyperparameters, and unbounded computational abilities (see Table 2). The realistic adversary is assumed to have no access to the original images and an imperfect reconstruction algorithm. Even these assumptions leave us with a rigorous threat model that may still be considered pessimistic for most real-world scenarios [1]. Having established the three relevant threat models to consider, Ziller, Mueller, Stieger, et al. compare AI model accuracy in conjunction with the reconstruction risk under each threat model at different values of ε. As we saw in Table 1, this is done for three exemplary Medical Imaging datasets. Their full results are presented in Table 3: Table 3: Comparison of AI model performance and reconstruction risk per threat model across the RadImageNet [7], HAM10000 [8], and MSD Liver [9] datasets with δ = 8⁻⁷⋅10 and various privacy budgets, including some as high as ε = 10⁹ and ε = 10¹². A higher MCC/Dice score indicates higher accuracy. Image by A. Ziller, T.T. Mueller, S. Stieger, et al from Table 3 in Reconciling privacy and accuracy in AI for medical imaging [1] (use under CC-BY 4.0 license). Unsurprisingly, high privacy budgets (exceeding ε = 10⁶) significantly mitigate the loss of accuracy seen with lower (stricter) privacy budgets. Across all tested datasets, models trained with high-budget DP at ε = 10⁹ (HAM10000, MSD Liver) or ε = 10¹² (RadImageNet) perform nearly as well as their non-privately trained counterparts. This is in line with our understanding of the privacy/accuracy tradeoff: the less noise introduced into the training data, the better a model can learn. What is surprising is the degree of empirical protection afforded by high-budget DP against reconstruction under the realistic threat model. Remarkably, the realistic reconstruction risk is assessed to be 0% for each of the aforementioned models. The high efficacy of high-budget DP in defending medical AI training images against realistic reconstruction attacks is made even clearer by looking at the results of reconstruction attempts. Figure 1 below shows the five most readily reconstructed images from the MSD Liver dataset [9] using DP with high privacy budgets of ε = 10⁶, ε = 10⁹, ε = 10¹², and ε = 10¹⁵. Figure 1: The five most readily reconstructed images from the MSD Liver dataset [9] using DP with high privacy budgets of ε = 10⁶, ε = 10⁹, ε = 10¹², and ε = 10¹⁵. Image by A. Ziller, T.T. Mueller, S. Stieger, et al from Figure 3 in Reconciling privacy and accuracy in AI for medical imaging [1] (use under CC-BY 4.0 license). Note that, at least to the naked eye, even the best reconstructions obtained when using the former two budgets are visually indistinguishable from random noise. This lends intuitive credence to the argument that budgets often deemed too high to provide any meaningful protection could be instrumental in protecting privacy without giving up accuracy when using AI for medical imaging. In contrast, the reconstructions when using ε = 10¹⁵ closely resemble the original images, showing that not all high budgets are created equal. Based on their findings, Ziller, Mueller, Stieger, et al. make the case for training medical imaging AI models using (at least) high-budget DP as the norm. They note the empirical efficacy of high-budget DP in countering realistic reconstruction risks at very little cost in terms of model accuracy. The authors go so far as to claim that “it seems negligent to train AI models without any form of formal privacy guarantee.” [1] Conclusion We started with a hypothetical scenario in which you were forced to decide between a bank vault door or no door at all for your dream home (or giving up and selling the incomplete house). After an exploration of the risks posed by inadequate privacy protection in medical AI, we looked into the privacy/accuracy tradeoff as well as the history and theory behind reconstruction attacks and differential privacy (DP). We then saw how DP with common privacy budgets (ε values) degrades medical AI model performance and compared it to the bank vault door in our hypothetical. Finally, we examined empirical results from the paper Reconciling privacy and accuracy in AI for medical imaging to find out how high-budget differential privacy can be used to escape the false dichotomy of bank vault door vs. no door and protect Patient Privacy in the real world without sacrificing model accuracy in the process. If you enjoyed this article, please consider following me on LinkedIn to keep up with future articles and projects. References [1] Ziller, A., Mueller, T.T., Stieger, S. et al. Reconciling privacy and accuracy in AI for medical imaging. Nat Mach Intell 6, 764–774 (2024). https://doi.org/10.1038/s42256-024-00858-y. [2] Ray, S. Samsung bans ChatGPT and other chatbots for employees after sensitive code leak. Forbes (2023). https://www.forbes.com/sites/siladityaray/2023/05/02/samsung-bans-chatgpt-and-other-chatbots-for-employees-after-sensitive-code-leak/. [3] Ateniese, G., Mancini, L. V., Spognardi, A. et al. Hacking smart machines with smarter ones: how to extract meaningful data from machine learning classifiers. International Journal of Security and Networks 10, 137–150 (2015). https://doi.org/10.48550/arXiv.1306.4447. [4] Dinur, I. & Nissim, K. Revealing information while preserving privacy. Proc. 22nd ACM SIGMOD-SIGACT-SIGART Symp Principles Database Syst 202–210 (2003). https://doi.org/10.1145/773153.773173. [5] Dwork, C. & Roth, A. The algorithmic foundations of differential privacy. Foundations and Trends in Theoretical Computer Science 9, 211–407 (2014). https://doi.org/10.1561/0400000042. [6] Dwork, C., Kohli, N. & Mulligan, D. Differential privacy in practice: expose your epsilons! Journal of Privacy and Confidentiality 9 (2019). https://doi.org/10.29012/jpc.689. [7] Mei, X., Liu, Z., Robson, P.M. et al. RadImageNet: an open radiologic deep learning research dataset for effective transfer learning. Radiol Artif Intell 4.5, e210315 (2022). https://doi.org/10.1148/ryai.210315. [8] Tschandl, P., Rosendahl, C. & Kittler, H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Sci Data 5, 180161 (2018). https://doi.org/10.1038/sdata.2018.161. [9] Antonelli, M., Reinke, A., Bakas, S. et al. The Medical Segmentation Decathlon. Nat Commun 13, 4128 (2022). https://doi.org/10.1038/s41467-022-30695-9.

Imagine you’re building your dream home. Just about everything is ready. All that’s left to do is pick out a front door. Since the neighborhood has a low crime rate, you decide you want a door with a standard lock — nothing too fancy, but probably enough to deter 99.9% of would-be burglars.

Unfortunately, the local homeowners’ association (HOA) has a rule stating that all front doors in the neighborhood must be bank vault doors. Their reasoning? Bank vault doors are the only doors that have been mathematically proven to be absolutely secure. As far as they’re concerned, any front door below that standard may as well not be there at all.

You’re left with three options, none of which seems particularly appealing:

Concede defeat and have a bank vault door installed. Not only is this expensive and cumbersome, but you’ll be left with a front door that bogs you down every single time you want to open or close it. At least burglars won’t be a problem!
Leave your house doorless. The HOA rule imposes requirements on any front door in the neighborhood, but it doesn’t technically forbid you from not installing a door at all. That would save you a lot of time and money. The downside, of course, is that it would allow anyone to come and go as they please. On top of that, the HOA could always close the loophole, taking you back to square one.
Opt out entirely. Faced with such a stark dilemma (all-in on either security or practicality), you choose not to play the game at all, selling your nearly-complete house and looking for someplace else to live.

This scenario is obviously completely unrealistic. In real life, everybody strives to strike an appropriate balance between security and practicality. This balance is informed by everyone’s own circumstances and risk analysis, but it universally lands somewhere between the two extremes of bank vault door and no door at all.

But what if instead of your dream home, you imagined a medical AI model that has the power to help doctors improve patient outcomes? Highly-sensitive training data points from patients are your valuables. The privacy protection measures you take are the front door you choose to install. Healthcare providers and the scientific community are the HOA.

Suddenly, the scenario is much closer to reality. In this article, we’ll explore why that is. After understanding the problem, we’ll consider a simple but empirically effective solution proposed in the paper Reconciling privacy and accuracy in AI for medical imaging [1]. The authors propose a balanced alternative to the three bad choices laid out above, much like the real-life approach of a typical front door.

The State of Patient Privacy in Medical AI

Over the past few years, artificial intelligence has become an ever more ubiquitous part of our day-to-day lives, proving its utility across a wide range of domains. The rising use of AI models has, however, raised questions and concerns about protecting the privacy of the data used to train them. You may remember the well-known case of ChatGPT, just months after its initial release, exposing proprietary code from Samsung [2].

Some of the privacy risks associated with AI models are obvious. For example, if the training data used for a model isn’t stored securely enough, bad actors could find ways to access it directly. Others are more insidious, such as the risk of reconstruction. As the name implies, in a reconstruction attack, a bad actor attempts to reconstruct a model’s training data without needing to gain direct access to the dataset.

Medical records are one of the most sensitive kinds of personal information there are. Although specific regulation varies by jurisdiction, patient data is generally subject to stringent safeguards, with hefty fines for inadequate protection. Beyond the letter of the law, unintentionally exposing such data could irreparably damage our ability to use specialized AI to empower medical professionals.

As Ziller, Mueller, Stieger, et al. point out [1], fully taking advantage of medical AI requires rich datasets comprising information from actual patients. This information must be obtained with the full consent of the patient. Ethically acquiring medical data for research was challenging enough as it was before the unique challenges posed by AI came into play. But if proprietary code being exposed caused Samsung to ban the use of ChatGPT [2], what would happen if attackers managed to reconstruct MRI scans and identify the patients they belonged to? Even isolated instances of negligent protection against data reconstruction could end up being a monumental setback for medical AI as a whole.

Tying this back into our front door metaphor, the HOA statute calling for bank vault doors starts to make a little bit more sense. When the cost of a single break-in could be so catastrophic for the entire neighborhood, it’s only natural to want to go to any lengths to prevent them.

Differential Privacy (DP) as a Theoretical Bank Vault Door

Before we discuss what an appropriate balance between privacy and practicality might look like in the context of medical AI, we have to turn our attention to the inherent tradeoff between protecting an AI model’s training data and optimizing for quality of performance. This will set the stage for us to develop a basic understanding of Differential Privacy (DP), the theoretical gold standard of privacy protection.

Although academic interest in training data privacy has increased significantly over the past four years, principles on which much of the conversation is based were pointed out by researchers well before the recent LLM boom, and even before OpenAI was founded in 2015. Though it doesn’t deal with reconstruction per se, the 2013 paper Hacking smart machines with smarter ones [3] demonstrates a generalizable attack methodology capable of accurately inferring statistical properties of machine learning classifiers, noting:

“Although ML algorithms are known and publicly released, training sets may not be reasonably ascertainable and, indeed, may be guarded as trade secrets. While much research has been performed about the privacy of the elements of training sets, […] we focus our attention on ML classifiers and on the statistical information that can be unconsciously or maliciously revealed from them. We show that it is possible to infer unexpected but useful information from ML classifiers.” [3]

Theoretical data reconstruction attacks were described even earlier, in a context not directly pertaining to machine learning. The landmark 2003 paper Revealing information while preserving privacy [4] demonstrates a polynomial-time reconstruction algorithm for statistical databases. (Such databases are intended to provide answers to questions about their data in aggregate while keeping individual data points anonymous.) The authors show that to mitigate the risk of reconstruction, a certain amount of noise needs to be introduced into the data. Needless to say, perturbing the original data in this way, while necessary for privacy, has implications for the quality of the responses to queries, i.e., the accuracy of the statistical database.

In explaining the purpose of DP in the first chapter of their book The Algorithmic Foundations of Differential Privacy [5], Cynthia Dwork and Aaron Roth address this tradeoff between privacy and accuracy:

“[T]he Fundamental Law of Information Recovery states that overly accurate answers to too many questions will destroy privacy in a spectacular way. The goal of algorithmic research on differential privacy is to postpone this inevitability as long as possible. Differential privacy addresses the paradox of learning nothing about an individual while learning useful information about a population.” [5]

The notion of “learning nothing about an individual while learning useful information about a population” is captured by considering two datasets that differ by a single entry (one that includes the entry and one that doesn’t). An (ε, δ)-differentially private querying mechanism is one for which the probability of a certain output being returned when querying one dataset is at most a multiplicative factor of the probability when querying the other dataset. Denoting the mechanism by M, the set of possible outputs by S, and the datasets by x and y, we formalize this as [5]:

Pr[M(x) ∈ S] ≤ exp(ε) ⋅ Pr[M(y) ∈ S] + δ

Where ε is the privacy loss parameter and δ is the failure probability parameter. ε quantifies how much privacy is lost as a result of a query, while a positive δ allows for privacy to fail altogether for a query at a certain (usually very low) probability. Note that ε is an exponential parameter, meaning that even slightly increasing it can cause privacy to decay significantly.

An important and useful property of DP is composition. Notice that the definition above only applies to cases where we run a single query. The composition property helps us generalize it to cover multiple queries based on the fact that privacy loss and failure probability accumulate predictably when we compose several queries, be they based on the same mechanism or different ones. This accumulation is easily proven to be (at most) linear [5]. What this means is that, rather than considering a privacy loss parameter for one query, we may view ε as a privacy budget that can be utilized across a number of queries. For example, when taken together, one query using a (1, 0)-DP mechanism and two queries using a (0.5, 0)-DP mechanism satisfy (2, 0)-DP.

The value of DP comes from the theoretical privacy guarantees it promises. Setting ε = 1 and δ = 0, for example, we find that the probability of any given output occurring when querying dataset y is at most exp(1) = e ≈ 2.718 times greater than that same output occurring when querying dataset x. Why does this matter? Because the greater the discrepancy between the probabilities of certain outputs occurring, the easier it is to determine the contribution of the individual entry by which the two datasets differ, and the easier it is to ultimately reconstruct that individual entry.

In practice, designing an (ε, δ)-differentially private randomized mechanism entails the addition of random noise drawn from a distribution dependent on ε and δ. The specifics are beyond the scope of this article. Shifting our focus back to machine learning, though, we find that the idea is the same: DP for ML hinges on introducing noise into the training data, which yields robust privacy guarantees in much the same way.

Of course, this is where the tradeoff we mentioned comes into play. Adding noise to the training data comes at the cost of making learning more difficult. We could absolutely add enough noise to achieve ε = 0.01 and δ = 0, making the difference in output probabilities between x and y virtually nonexistent. This would be wonderful for privacy, but terrible for learning. A model trained on such a noisy dataset would perform very poorly on most tasks.

There is no consensus on what constitutes a “good” ε value, or on universal methodologies or best practices for ε selection [6]. In many ways, ε embodies the privacy/accuracy tradeoff, and the “proper” value to aim for is highly context-dependent. ε = 1 is generally regarded as offering high privacy guarantees. Although privacy diminishes exponentially with respect to ε, values as high as ε = 32 are mentioned in literature and thought to provide moderately strong privacy guarantees [1].

The authors of Reconciling privacy and accuracy in AI for medical imaging [1] test the effects of DP on the accuracy of AI models on three real-world medical imaging datasets. They do so using various values of ε and comparing them to a non-private (non-DP) control. Table 1 provides a partial summary of their results for ε = 1 and ε = 8:

**Table 1:** Comparison of AI model performance across the RadImageNet [7], HAM10000 [8], and MSD Liver [9] datasets with *δ = 8*⁻⁷⋅10 and privacy budgets of ε = 1, ε = 8, and without DP (non-private). A higher MCC/Dice score indicates higher accuracy. Although providing strong theoretical privacy guarantees in the face of a worst-case adversary, **DP significantly degrades model accuracy.** The negative impact on performance is especially noticeable in the latter two datasets, which are considered small datasets. Image by the author, based on image by A. Ziller, T.T. Mueller, S. Stieger, et al from Table 3 in ***Reconciling privacy and accuracy in AI for medical imaging*** [1] (use under CC-BY 4.0 license).

Even approaching the higher end of the typical ε values attested in literature, DP is still as cumbersome as a bank vault door for medical imaging tasks. The noise introduced into the training data is catastrophic for AI model accuracy, especially when the datasets at hand are small. Note, for example, the huge drop-off in Dice score on the MSD Liver dataset, even with the relatively high ε value of 8.

Ziller, Mueller, Stieger, et al. suggest that the accuracy drawbacks of DP with typical ε values may contribute to the lack of widespread adoption of DP in the field of Medical Ai [1]. Yes, wanting mathematically-provable privacy guarantees is definitely sensible, but at what cost? Leaving so much of the diagnostic power of AI models on the table in the name of privacy is not an easy choice to make.

Revisiting our dream home scenario armed with an understanding of DP, we find that the options we (seem to) have map neatly onto the three we had for our front door.

DP with typical values of ε is like installing a bank vault door: costly, but effective for privacy. As we’ll see, it’s also complete overkill in this case.
Not using DP is like not installing a door at all: much easier, but risky. As mentioned above, though, DP has yet to be widely applied in medical AI [1].
Passing up opportunities to use AI is like giving up and selling the house: it saves us the headache of dealing with privacy concerns weighed against incentives to maximize accuracy, but a lot of potential is lost in the process.

It looks like we’re at an impasse… unless we think outside the box.

High-Budget DP: Privacy and Accuracy Aren’t an Either/Or

In Reconciling privacy and accuracy in AI for medical imaging [1], Ziller, Mueller, Stieger, et al. offer the medical AI equivalent of a regular front door — an approach that manages to protect privacy while giving up very little in the way of model performance. Granted, this protection is not theoretically optimal — far from it. However, as the authors show through a series of experiments, it is good enough to counter almost any realistic threat of reconstruction.

As the saying goes, “Perfect is the enemy of good.” In this case, it is the “optimal” — an insistence on arbitrarily low ε values — that locks us into the false dichotomy of total privacy versus total accuracy. Just as a bank vault door has its place in the real world, so does DP with ε ≤ 32. Still, the existence of the bank vault door doesn’t mean plain old front doors don’t also have a place in the world. The same goes for high-budget DP.

The idea behind high-budget DP is straightforward: using privacy budgets (ε values) that are so high that they “are near-universally shunned as being meaningless” [1] — budgets ranging from ε = 10⁶ to as high as ε = 10¹⁵. In theory, these provide such weak privacy guarantees that it seems like common sense to dismiss them as no better than not using DP at all. In practice, though, this couldn’t be further from the truth. As we will see by looking at the results from the paper, high-budget DP shows significant promise in countering realistic threats. As Ziller, Mueller, Stieger, et al. put it [1]:

“[E]ven a ‘pinch of privacy’ has drastic effects in practical scenarios.”

First, though, we need to ask ourselves what we consider to be a “realistic” threat. Any discussion of the efficacy of high-budget DP is inextricably tied to the threat model under which we choose to evaluate it. In this context, a threat model is simply the set of assumptions we make about what a bad actor interested in obtaining our model’s training data is able to do.

**Table 2:** Comparison of threat models. For all three, we also assume that the adversary has unbounded computational ability. Image by A. Ziller, T.T. Mueller, S. Stieger, et al from Table 1 in ***Reconciling privacy and accuracy in AI for medical imaging*** [1] (use under CC-BY 4.0 license).

The paper’s findings hinge on a calibration of the assumptions to better suit real-world threats to patient privacy. The authors argue that the worst-case model, which is the one typically used for DP, is far too pessimistic. For example, it assumes that the adversary has full access to each original image while attempting to reconstruct it based on the AI model (see Table 2) [1]. This pessimism explains the discrepancy between the reported “drastic effects in practical scenarios” of high privacy budgets and the very weak theoretical privacy guarantees that they offer. We may liken it to incorrectly assessing the security threats a typical house faces, wrongly assuming they are likely to be as sophisticated and enduring as those faced by a bank.

The authors therefore propose two alternative threat models, which they call the “relaxed” and “realistic” models. Under both of these, adversaries keep some core capabilities from the worst-case model: access to the AI model’s architecture and weights, the ability to manipulate its hyperparameters, and unbounded computational abilities (see Table 2). The realistic adversary is assumed to have no access to the original images and an imperfect reconstruction algorithm. Even these assumptions leave us with a rigorous threat model that may still be considered pessimistic for most real-world scenarios [1].

Having established the three relevant threat models to consider, Ziller, Mueller, Stieger, et al. compare AI model accuracy in conjunction with the reconstruction risk under each threat model at different values of ε. As we saw in Table 1, this is done for three exemplary Medical Imaging datasets. Their full results are presented in Table 3:

**Table 3:** Comparison of AI model performance and reconstruction risk per threat model across the RadImageNet [7], HAM10000 [8], and MSD Liver [9] datasets with *δ = 8*⁻⁷⋅10 and various privacy budgets, including some as high as ε = 10⁹ and ε = 10¹². A higher MCC/Dice score indicates higher accuracy. Image by A. Ziller, T.T. Mueller, S. Stieger, et al from Table 3 in ***Reconciling privacy and accuracy in AI for medical imaging*** [1] (use under CC-BY 4.0 license).

Unsurprisingly, high privacy budgets (exceeding ε = 10⁶) significantly mitigate the loss of accuracy seen with lower (stricter) privacy budgets. Across all tested datasets, models trained with high-budget DP at ε = 10⁹ (HAM10000, MSD Liver) or ε = 10¹² (RadImageNet) perform nearly as well as their non-privately trained counterparts. This is in line with our understanding of the privacy/accuracy tradeoff: the less noise introduced into the training data, the better a model can learn.

What is surprising is the degree of empirical protection afforded by high-budget DP against reconstruction under the realistic threat model. Remarkably, the realistic reconstruction risk is assessed to be 0% for each of the aforementioned models. The high efficacy of high-budget DP in defending medical AI training images against realistic reconstruction attacks is made even clearer by looking at the results of reconstruction attempts. Figure 1 below shows the five most readily reconstructed images from the MSD Liver dataset [9] using DP with high privacy budgets of ε = 10⁶, ε = 10⁹, ε = 10¹², and ε = 10¹⁵.

Figure 1: The five most readily reconstructed images from the MSD Liver dataset [9] using DP with high privacy budgets of ε = 10⁶, ε = 10⁹, ε = 10¹², and ε = 10¹⁵. Image by A. Ziller, T.T. Mueller, S. Stieger, et al from Figure 3 in Reconciling privacy and accuracy in AI for medical imaging [1] (use under CC-BY 4.0 license).

Note that, at least to the naked eye, even the best reconstructions obtained when using the former two budgets are visually indistinguishable from random noise. This lends intuitive credence to the argument that budgets often deemed too high to provide any meaningful protection could be instrumental in protecting privacy without giving up accuracy when using AI for medical imaging. In contrast, the reconstructions when using ε = 10¹⁵ closely resemble the original images, showing that not all high budgets are created equal.

Based on their findings, Ziller, Mueller, Stieger, et al. make the case for training medical imaging AI models using (at least) high-budget DP as the norm. They note the empirical efficacy of high-budget DP in countering realistic reconstruction risks at very little cost in terms of model accuracy. The authors go so far as to claim that “it seems negligent to train AI models without any form of formal privacy guarantee.” [1]

Conclusion

We started with a hypothetical scenario in which you were forced to decide between a bank vault door or no door at all for your dream home (or giving up and selling the incomplete house). After an exploration of the risks posed by inadequate privacy protection in medical AI, we looked into the privacy/accuracy tradeoff as well as the history and theory behind reconstruction attacks and differential privacy (DP). We then saw how DP with common privacy budgets (ε values) degrades medical AI model performance and compared it to the bank vault door in our hypothetical.

Finally, we examined empirical results from the paper Reconciling privacy and accuracy in AI for medical imaging to find out how high-budget differential privacy can be used to escape the false dichotomy of bank vault door vs. no door and protect Patient Privacy in the real world without sacrificing model accuracy in the process.

If you enjoyed this article, please consider following me on LinkedIn to keep up with future articles and projects.

References

[1] Ziller, A., Mueller, T.T., Stieger, S. et al. Reconciling privacy and accuracy in AI for medical imaging. Nat Mach Intell 6, 764–774 (2024). https://doi.org/10.1038/s42256-024-00858-y.

[2] Ray, S. Samsung bans ChatGPT and other chatbots for employees after sensitive code leak. Forbes (2023). https://www.forbes.com/sites/siladityaray/2023/05/02/samsung-bans-chatgpt-and-other-chatbots-for-employees-after-sensitive-code-leak/.

[3] Ateniese, G., Mancini, L. V., Spognardi, A. et al. Hacking smart machines with smarter ones: how to extract meaningful data from machine learning classifiers. International Journal of Security and Networks 10, 137–150 (2015). https://doi.org/10.48550/arXiv.1306.4447.

[4] Dinur, I. & Nissim, K. Revealing information while preserving privacy. Proc. 22nd ACM SIGMOD-SIGACT-SIGART Symp Principles Database Syst 202–210 (2003). https://doi.org/10.1145/773153.773173.

[5] Dwork, C. & Roth, A. The algorithmic foundations of differential privacy. Foundations and Trends in Theoretical Computer Science 9, 211–407 (2014). https://doi.org/10.1561/0400000042.

[6] Dwork, C., Kohli, N. & Mulligan, D. Differential privacy in practice: expose your epsilons! Journal of Privacy and Confidentiality 9 (2019). https://doi.org/10.29012/jpc.689.

[7] Mei, X., Liu, Z., Robson, P.M. et al. RadImageNet: an open radiologic deep learning research dataset for effective transfer learning. Radiol Artif Intell 4.5, e210315 (2022). https://doi.org/10.1148/ryai.210315.

[8] Tschandl, P., Rosendahl, C. & Kittler, H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Sci Data 5, 180161 (2018). https://doi.org/10.1038/sdata.2018.161.

[9] Antonelli, M., Reinke, A., Bakas, S. et al. The Medical Segmentation Decathlon. Nat Commun 13, 4128 (2022). https://doi.org/10.1038/s41467-022-30695-9.

Stay Ahead

Explore More Insights

Stay ahead with more perspectives on cutting-edge power, infrastructure, energy, bitcoin and AI solutions. Explore these articles to uncover strategies and insights shaping the future of industries.

IBM targets AI application growth with DataStax buy

In particular IBM said DataStax’s technology will be built into its watsonx portfolio of generative AI products to help manage the vast amounts of unstructured data used in generative AI application development. Thousands of organizations including FedEx, Capital One, The Home Depot and Verizon use Apache Cassandra, and it offers

New Relic boosts observability platform with AI intelligence

New Relic announced updates to its Intelligent Observability Platform this week, which the company says will enable a unified view of system relationships and dependencies and intelligently connect technical systems with business context. New Relic’s cloud-based observability platform monitors applications and services in real time to provide insights into software,

Alibaba to invest $52 billion in AI and cloud over three years

“Although $52 billion is a lot of money, every US-based hyper scaler is committing more than this to AI already,” said Hyoun Park, CEO and chief analyst at Amalgam Insights. “AWS is committing roughly $100 billion to new capital expenditures next year, most of which is servers for AI-related use

Intel targets edge, high-performance computing with extended Xeon 6 chip line

“As we grow the core counts, do people really need to have a four socket system or an eight socket system? And I think the answer very simply is yes, we continue to see a strong demand for people who want to scale out their memory capacity,” said Ronak Singhal,

Oil Slumps as US Confidence Dives

Oil slumped along with equity markets as US consumer confidence tumbled, adding to mounting concerns that US President Donald Trump’s policies will hamper economic growth and sap energy demand. West Texas Intermediate fell 2.5% to settle below $69 a barrel at the lowest closing price this year. US consumer confidence dropped the most since 2021 and missed analysts’ estimates, prompting traders to flee risk assets, including equities. Trump’s tariffs and recent moves to further decouple economic ties with China, which spurred a drop in the Asian country’s stock markets Tuesday, are worsening the already-gloomy outlook for energy demand in the world’s largest oil consumer. Domestically, the trade turmoil is raising Americans’ inflation expectations amid a cooling labor market. “Crude markets are seeing another layer of bearish pressure from a continued string of misses in economic data,” said Frank Monkam, head of macro trading at Buffalo Bayou Commodities. “Such a rollover in economic data bodes ill for crude demand.” Crude has now broken below the roughly $5 range it had wandered in for February. Oil had initially spiked above $80 early this year before fading amid persistent expectations of lackluster Chinese demand, the potential for additional barrels on the market and the prospect that tariffs will weigh on global growth. Earlier this week, the US imposed more curbs on brokers, vessels and individuals that it said were linked to illicit shipments of Iranian crude. Markets had a muted reaction to the additional sanctions on expectations that the trade would adapt quickly by ramping up ship-to-ship transfers or switching off geo-locating signals for longer. The shifts would resemble Russia’s steps to keep crude exports flowing in the face of restrictions. “Sanctions are not the bullish factor many are expecting unless we see true attempts at locating and blockading tankers with naval

Iran Rejects Direct Nuclear Talks With USA Under Trump Policy

Iran’s Foreign Minister Abbas Araghchi said his country won’t agree to direct nuclear talks with the US while President Donald Trump persists with his hard-line policy against the Islamic Republic. “We will not negotiate under pressure, sanctions, or threats,” Araghchi said in a televised press conference alongside his Russian counterpart Sergei Lavrov in Tehran on Tuesday. Direct negotiations between Iran and the US on the nuclear issue will be impossible “as long as maximum pressure is being applied in this manner,” Araghchi said. Trump has vowed to squeeze Iran’s economy and target its oil exports as part of a return to the “maximum pressure” strategy that he deployed in his first term. That led to the 2018 US withdrawal from a landmark international deal that limited Iran’s atomic activities in exchange for sanctions relief. Iran’s stockpile of near weapons-grade enriched uranium has since surged. Araghchi said Iran will cooperate fully on its nuclear affairs — but only with “friends” China and Russia. Since returning to office last month, Trump has said he wants Iran to agree to a new nuclear agreement, but Iran’s Supreme Leader Ayatollah Ali Khamenei dismissed the idea earlier this month, saying negotiating with the US “won’t solve any of the country’s problems.” Speaking alongside Araghchi, Lavrov said Russia will pursue diplomatic efforts to resolve the Iranian nuclear issue. “We’re convinced that the tool of diplomacy remains,” Lavrov said. “It cannot be neglected — it must be used as effectively as possible without any threats and without hints at the possibility of certain forceful solutions.” Iran and the US haven’t had direct, formal ties since the 1979 Islamic revolution and previous negotiations that led to the 2015 nuclear deal took place through mediators. Araghchi didn’t mention whether indirect or mediated talks with the US are still on

Schneider Electric books strong 2024 revenue, earnings growth amid data center boom

Dive Brief: Schneider Electric saw stronger-than-forecast revenue and adjusted earnings in 2024 as its customers’ data center investments — especially in North America — drove mid-double-digit growth in its energy management business, the company said Feb. 20. Year-over-year organic revenue growth accelerated 12% in the fourth quarter, supported by 25% growth in Schneider Electric’s North American energy management business. The company’s year-end 2024 sales backlog of 21.4 billion euros, or about $22.39 billion, was its highest ever, and the company plans to invest about 2 billion euros through 2027 to expand production capacity, it said. Much of the planned capacity expansion will occur in North America despite uncertainty around U.S. trade policy that could necessitate “commercial actions” to protect the company’s profitability, Chief Financial Officer Hilary Maxson said Thursday on Schneider Electric’s earnings call. Dive Insight: Schneider data centers and networks end-market has been strong throughout 2024 and should continue to see robust growth in 2025 and beyond, CEO Olivier Blum said on his first earnings call since replacing Peter Herweck in November. The AI investment boom supports annual growth of 10% or more through 2027 in the company’s data centers and networks business, which accounts for 24% of Schneider Electric’s 2024 end-market exposure, the company said in its earnings presentation. “Pure data centers” make up 20% of Schneider Electric’s end market exposure, with hyperscalers contributing “a bit less than half” of that total, Maxson said. “Suffice to say we feel there is healthy growth in that segment … and we believe there is healthy growth to come, [though] not exponential … as this new infrastructure backbone is built out,” Maxson said. Schneider expects DeepSeek, a Chinese AI firm that caught the industry off-guard in January when it released a reasoning model that appeared to use far less energy than

New York PSC approves retail and residential storage plan as 6-GW 2030 target in question

Dive Brief: The New York State Public Service Commission has approved the state’s retail and residential energy storage implementation plan, a significant step in its effort to reach 6 GW of energy storage by 2030. The Feb. 13 order approved a framework to reach the state’s retail storage deployment goal of 1,500 MW and its residential storage deployment goal of 200 MW. It also includes incentives for resources participating in the New York Independent System Operator’s distributed energy resources program to also be eligible for the retail storage incentive, the PSC said. The plan was approved as a new forecast by Aurora Energy Research shows New York falling “marginally short” of its 2030 energy storage target despite an expected deployment surge in the late 2020s, but reaching 30 GW of deployed storage capacity by 2050. Dive Insight: New York’s 6-GW 2030 goal will “support a buildout of storage deployments estimated to reduce projected future statewide electric system costs by nearly $2 billion, in addition to further benefits in the form of improved public health because of reduced exposure to harmful fossil fuel pollutants,” the PSC said in announcing the order. The 6-GW goal represents a doubling of the previous 2030 goal of 3 GW. It envisions 1.7 GW of new retail and residential storage plus 3 GW of new bulk storage added to about 1.3 GW of existing storage assets being procured by or under contract with the state as of April 1, 2024, the PSC said on Feb. 13. Following the adoption this month of its retail and residential implementation plan, the New York State Energy Research and Development Authority expects to make the first of three annual bulk storage solicitations by the end of June for deployment in 2027 and 2028. It plans subsequent storage solicitations in 2026

Charging Forward: UK battery storage projects reach startup, grid delays and more

In this week’s Charging Forward, Gore Street, Eku and BW ESS reach energisation at UK battery energy storage system (BESS) projects, amid warnings over an oversubscribed grid connection queue. This week’s headlines: Root-Power secures planning consent for 40 MW Rotherham BESS Sungrow and BW ESS Bramley Project begins operations Warnings over UK grid connection queue Invinity and Frontier Power partner on UK long duration energy storage projects Fire at Statera BESS site in Essex brought under control Gore Street energises UK Enderby BESS project Eku energises two UK BESS projects International news: China and Saudi Arabia collaborate on 12.5 GWh of energy storage projects and Canadian firm Hydrostor secures $200 million for compressed air energy storage Root-Power consent for 40 MW Rotherham BESS UK energy storage developer Root-Power has secured planning consent for a 40 MW/80 MWh BESS project in Brinsworth, Rotherham. Root-Power said the site will power 80,000 homes for a two-hour period once fully operational, and delivering a biodiversity net gain of 32.76%. The Brinsworth BESS is the fourth planning approval for Root-Power in 2025, following consents at sites in Yorkshire, County Durham and the Scottish Highlands. © Supplied by Root-PowerThe site of Root-Power’s 40 MW/80 MWh Brinsworth BESS project in Rotherham. Root-Power managing director Neil Brooks said the company “carefully selected a near perfect location” for the Brinsworth project. “Managing competing constraints is always difficult when planning a project, so finding a suitable location only 1 mile from the point of connection in an urban area, without causing unacceptable noise or visual impact on sensitive receptors is a real achievement,” he said. “We are happy to see that the planning committee unanimously supported our application, which is a real vote of confidence in our process and team.” Sungrow and BW ESS Bramley BESS starts up Swiss energy storage developer

Costain secures multi-million pound Sizewell C contract

UK construction and engineering firm Costain (LON:COST) has secured a multi-million pound contract to support the construction of the Sizewell C nuclear power plant. Costain said under the ten-year framework agreement, the company will provide support in areas such as delivery integration, health and safety and quality control. French state-owned energy firm EDF is developing the 3.2 GW nuclear power station, which could provide up to 7% of UK energy needs over its 60-year lifetime. The UK government holds a 76.1% stake in Sizewell C, with EDF holding the remaining 23.9%. Costain defence and nuclear energy sector director Bob Anstey said the Sizewell C project is a “vital part of creating a sustainable future”. “We have a long and successful track record in delivering for our civil nuclear customers, with a highly qualified and experienced workforce that consistently works to the highest safety and quality standards,” Anstey said. “A key part of our role will be to help ensure the project leaves a positive legacy, and we look forward to working closely with Sizewell C on a range of social value and employment initiatives that improve lives and provide long-term benefits to local communities.” Sizewell C Ltd managing director Nigel Cann said the project will “strengthen energy security and provide clean, reliable electricity for millions”. “We welcome Costain to the Sizewell C supplier family,” Cann said. “We are committed to providing thousands of great jobs and career development opportunities and we’re looking forward to working with our suppliers to boost skills, promote a diverse workforce and spread opportunities as widely as possible.” Sizewell C criticism The Sizewell C project has attracted significant criticism amid concerns over its ballooning costs. Earlier this year, campaign group Together Against Sizewell C (TASC) wrote to the National Audit Office calling for a review of

Cisco, Nvidia expand AI partnership to include Silicon One technology

In addition, Cisco and Nvidia will invest in cross-portfolio technology to tackle common challenges like congestion management and load balancing, ensuring that enterprises can accelerate their AI deployments, Patel stated. The vendors said they would also collaborate to create and validate Nvidia Cloud Partner (NCP) and Enterprise Reference Architectures based on Nvidia Spectrum-X with Cisco Silicon One, Hyperfabric, Nexus, UCS Compute, Optics, and other Cisco technologies. History of Cisco, Nvidia collaborations The announcement is just the latest expansion of the Cisco/Nvidia partnership. The companies have already worked together to make Nvidia’s Tensor Core GPUs available in Cisco’s Unified Computing System (UCS) rack and blade servers, including Cisco UCS X-Series and UCS X-Series Direct, to support AI and data-intensive workloads in the data center and at the edge. The integrated package includes Nvidia AI Enterprise software, which features pretrained models and development tools for production-ready AI. Earlier this month, Cisco said it has shipped the UCS C845A M8 Rack Server for enterprise data center environments. The 8U rack server is built on Nvidia’s HGX platform and designed to deliver the accelerated compute capabilities needed for AI workloads such as LLM training, model fine-tuning, large model inferencing, and retrieval-augmented generation (RAG). The companies are also collaborating on AI Pods, which are preconfigured, validated, and optimized infrastructure packages that customers can plug into their data center or edge environments as needed. The Pods are based on Cisco Validated Design principals, which provide a blueprint for building reliable, scalable, and secure network infrastructures, according to Cisco. The Pods include Nvidia AI Enterprise, which features pretrained models and development tools for production-ready AI, and are managed through Cisco Intersight.

3 strategies for carbon-free data centers

Because of the strain that data centers (as well as other electrification sources, such as electric vehicles) are putting on the grid, “the data center industry needs to develop new power supply strategies to support growth plans,” Dietrich said. Here are the underling factors that play into the three strategies outlined by Uptime. Scale creates new opportunities: It’s not just that more data centers are being built, but the data centers under construction are fundamentally different in terms of sheer magnitude. For example, a typical enterprise data center might require between 10 and 25 megawatts of power. Today, the hyperscalers are building data centers in the 250-megawatt range and a large data center campus could require 1,000 megawatts of power. Data centers not only require a reliable source of power, they also require backup power in the form of generators. Dietrich pointed out that if a data center operator builds out enough backup capacity to support 250 megawatts of demand, they’re essentially building a new, on-site power plant. On the one hand, that new power plant requires permitting, it’s costly, and it requires highly training staffers to operate. On the other hand, it provides an opportunity. Instead of letting this asset sit around unused except in an emergency, organizations can leverage these power plants to generate energy that can be sold back to the grid. Dietrich described this arrangement as a win-win that enables the data center to generate revenue, and it helps the utility to gain a new source of power. Realistic expectations: Alternative energy sources like wind and solar, which are dependent on environmental factors, can’t technically or economically supply 100% of data center power, but they can provide a significant percentage of it. Organizations need to temper their expectations, Dietrich said.

Questions arise about reasons why Microsoft has cancelled data center lease plans

This, the company said, “allows us to invest and allocate resources to growth areas for our future. Our plans to spend over $80 billion on infrastructure this fiscal year remains on track as we continue to grow at a record pace to meet customer demand.” When asked for his reaction to the findings, John Annand, infrastructure and operations research practice lead at Info-Tech Research Group, pointed to a blog released last month by Microsoft president Brad Smith, and said he thinks the company “is hedging its bets. It reaffirms the $80 billion AI investment guidance in 2025, $40 billion in the US. Why lease when you can build/buy your own?” Over the past four years, he said, Microsoft “has been leasing more data centers than owning. Perhaps they are using the fact that the lessors are behind schedule on providing facilities or the power upgrades required to bring that ratio back into balance. The limiting factor for data centers has always been the availability of power, and this has only become more true with power-hungry AI workloads.” The company, said Annand, “has made very public statements about owning nuclear power plants to help address this demand. If third-party data center operators are finding it tough to provide Microsoft with the power they need, it would make sense that Microsoft vertically integrate its supply chain; so, cancel leases or statements of qualification in favor of investing in the building of their own capacity.” However, Gartner analyst Tony Harvey said of the report, “so much of this is still speculation.” Microsoft, he added, “has not stated as yet that they are reducing their capex spend, and there are reports that Microsoft have strongly refuted that they are making changes to their data center strategy.” The company, he said, “like any other hyperscaler,

Quantum Computing Advancements Leap Forward In Evolving Data Center and AI Landscape

Overcoming the Barriers to Quantum Adoption Despite the promise of quantum computing, widespread deployment faces multiple hurdles: High Capital Costs: Quantum computing infrastructure requires substantial investment, with uncertain return-on-investment models. The partnership will explore cost-sharing strategies to mitigate risk. Undefined Revenue Models: Business frameworks for quantum services, including pricing structures and access models, remain in development. Hardware Limitations: Current quantum processors still struggle with error rates and scalability, requiring advancements in error correction and hybrid computing approaches. Software Maturity: Effective algorithms for leveraging quantum computing’s advantages remain an active area of research, particularly in real-world AI and optimization problems. SoftBank’s strategy includes leveraging its extensive telecom infrastructure and AI expertise to create real-world testing environments for quantum applications. By integrating quantum into existing data center operations, SoftBank aims to position itself at the forefront of the quantum-AI revolution. A Broader Play in Advanced Computing SoftBank’s quantum initiative follows a series of high-profile moves into the next generation of computing infrastructure. The company has been investing heavily in AI data centers, aligning with its “Beyond Carrier” strategy that expands its focus beyond telecommunications. Recent efforts include the development of large-scale AI models tailored to Japan and the enhancement of radio access networks (AI-RAN) through AI-driven optimizations. Internationally, SoftBank has explored data center expansion opportunities beyond Japan, as part of its efforts to support AI, cloud computing, and now quantum applications. The company’s long-term vision suggests that quantum data centers could eventually play a role in supporting AI-driven workloads at scale, offering performance benefits that classical supercomputers cannot achieve. The Road Ahead SoftBank and Quantinuum’s collaboration signals growing momentum for quantum computing in enterprise settings. While quantum remains a long-term bet, integrating QPUs into data center infrastructure represents a forward-looking approach that could redefine high-performance computing in the years to come. With

STACK Infrastructure Pushes Aggressive Data Center Expansion and Sustainability Strategy Into 2025

Global data center developer and operator STACK Infrastructure is providing a growing range of digital infrastructure solutions for hyperscalers, cloud service providers, and enterprise clients. Like almost all of the cutting-edge developers in the industry, Stack is maintaining the focus on scalability, reliability, and sustainability while delivering a full range of solutions, including build-to-suit, colocation, and powered shell facilities, with continued development in key global markets. Headquartered in the United States, the company has expanded its presence across North America, Europe, and Asia-Pacific, catering to the increasing demand for high-performance computing, artificial intelligence (AI), and cloud-based workloads. The company is known for its commitment to sustainable growth, leveraging green financing initiatives, energy-efficient designs, and renewable power sources to minimize its environmental impact. Through rapid expansion in technology hubs like Silicon Valley, Northern Virginia, Malaysia, and Loudoun County, the company continues to develop industry benchmarks for innovation and infrastructure resilience. With a customer-centric approach and a robust development pipeline, STACK Infrastructure is shaping the future of digital connectivity and data management in an era of accelerating digital transformation. Significant Developments Across 23 Major Data Center Markets Early in 2024, Stack broke ground on the expansion of their existing 100 MW campus in San Jose, servicing the power constrained Silicon Valley. Stack worked with the city of San Jose to add a 60 MW expansion to their SVY01 data center. While possibly the highest profile of Stack’s developments, due to its location, at that point in time the company had announced significant developments across 23 major data center markets, including: Stack’s 48 MW Santa Clara data center, featuring immediately available shell space powered by an onsite substation with rare, contracted capacity. Stack’s 56 MW Toronto campus, spanning 19 acres, includes an existing 8 MW data center and 48 MW expansion capacity,

Meta Update: Opens Mesa, Arizona Data Center; Unveils Major Subsea Cable Initiative; Forges Oklahoma Wind Farm PPA; More

Meta’s Project Waterworth: Building the Global Backbone for AI-Powered Digital Infrastructure Also very recently, Meta unveiled its most ambitious subsea cable initiative yet: Project Waterworth. Aimed at revolutionizing global digital connectivity, the project will span over 50,000 kilometers—surpassing the Earth’s circumference—and connect five major continents. When completed, it will be the world’s longest subsea cable system, featuring the highest-capacity technology available today. A Strategic Expansion to Key Global Markets As announced on Feb. 14, Project Waterworth is designed to enhance connectivity across critical regions, including the United States, India, Brazil, and South Africa. These regions are increasingly pivotal to global digital growth, and the new subsea infrastructure will fuel economic cooperation, promote digital inclusion, and unlock opportunities for technological advancement. In India, for instance, where rapid digital infrastructure growth is already underway, the project will accelerate progress and support the country’s ambitions for an expanded digital economy. This enhanced connectivity will foster regional integration and bolster the foundation for next-generation applications, including AI-driven services. Strengthening Global Digital Highways Subsea cables are the unsung heroes of global digital infrastructure, facilitating over 95% of intercontinental data traffic. With a multi-billion-dollar investment, Meta aims to open three new oceanic corridors that will deliver the high-speed, high-capacity bandwidth needed to fuel innovations like artificial intelligence. Meta’s experience in subsea infrastructure is extensive. Over the past decade, the company has collaborated with various partners to develop more than 20 subsea cables, including systems boasting up to 24 fiber pairs—far exceeding the typical 8 to 16 fiber pairs found in most new deployments. This technological edge ensures scalability and reliability, essential for handling the world’s ever-increasing data demands. Engineering Innovations for Resilience and Capacity Project Waterworth isn’t just about scale—it’s about resilience and cutting-edge engineering. The system will be the longest 24-fiber-pair subsea cable ever built, enhancing

Microsoft will invest $80B in AI data centers in fiscal 2025

And Microsoft isn’t the only one that is ramping up its investments into AI-enabled data centers. Rival cloud service providers are all investing in either upgrading or opening new data centers to capture a larger chunk of business from developers and users of large language models (LLMs). In a report published in October 2024, Bloomberg Intelligence estimated that demand for generative AI would push Microsoft, AWS, Google, Oracle, Meta, and Apple would between them devote $200 billion to capex in 2025, up from $110 billion in 2023. Microsoft is one of the biggest spenders, followed closely by Google and AWS, Bloomberg Intelligence said. Its estimate of Microsoft’s capital spending on AI, at $62.4 billion for calendar 2025, is lower than Smith’s claim that the company will invest $80 billion in the fiscal year to June 30, 2025. Both figures, though, are way higher than Microsoft’s 2020 capital expenditure of “just” $17.6 billion. The majority of the increased spending is tied to cloud services and the expansion of AI infrastructure needed to provide compute capacity for OpenAI workloads. Separately, last October Amazon CEO Andy Jassy said his company planned total capex spend of $75 billion in 2024 and even more in 2025, with much of it going to AWS, its cloud computing division.

John Deere unveils more autonomous farm machines to address skill labor shortage

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Self-driving tractors might be the path to self-driving cars. John Deere has revealed a new line of autonomous machines and tech across agriculture, construction and commercial landscaping. The Moline, Illinois-based John Deere has been in business for 187 years, yet it’s been a regular as a non-tech company showing off technology at the big tech trade show in Las Vegas and is back at CES 2025 with more autonomous tractors and other vehicles. This is not something we usually cover, but John Deere has a lot of data that is interesting in the big picture of tech. The message from the company is that there aren’t enough skilled farm laborers to do the work that its customers need. It’s been a challenge for most of the last two decades, said Jahmy Hindman, CTO at John Deere, in a briefing. Much of the tech will come this fall and after that. He noted that the average farmer in the U.S. is over 58 and works 12 to 18 hours a day to grow food for us. And he said the American Farm Bureau Federation estimates there are roughly 2.4 million farm jobs that need to be filled annually; and the agricultural work force continues to shrink. (This is my hint to the anti-immigration crowd). John Deere’s autonomous 9RX Tractor. Farmers can oversee it using an app. While each of these industries experiences their own set of challenges, a commonality across all is skilled labor availability. In construction, about 80% percent of contractors struggle to find skilled labor. And in commercial landscaping, 86% of landscaping business owners can’t find labor to fill open positions, he said. “They have to figure out how to do

2025 playbook for enterprise AI success, from agents to evals

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More 2025 is poised to be a pivotal year for enterprise AI. The past year has seen rapid innovation, and this year will see the same. This has made it more critical than ever to revisit your AI strategy to stay competitive and create value for your customers. From scaling AI agents to optimizing costs, here are the five critical areas enterprises should prioritize for their AI strategy this year. 1. Agents: the next generation of automation AI agents are no longer theoretical. In 2025, they’re indispensable tools for enterprises looking to streamline operations and enhance customer interactions. Unlike traditional software, agents powered by large language models (LLMs) can make nuanced decisions, navigate complex multi-step tasks, and integrate seamlessly with tools and APIs. At the start of 2024, agents were not ready for prime time, making frustrating mistakes like hallucinating URLs. They started getting better as frontier large language models themselves improved. “Let me put it this way,” said Sam Witteveen, cofounder of Red Dragon, a company that develops agents for companies, and that recently reviewed the 48 agents it built last year. “Interestingly, the ones that we built at the start of the year, a lot of those worked way better at the end of the year just because the models got better.” Witteveen shared this in the video podcast we filmed to discuss these five big trends in detail. Models are getting better and hallucinating less, and they’re also being trained to do agentic tasks. Another feature that the model providers are researching is a way to use the LLM as a judge, and as models get cheaper (something we’ll cover below), companies can use three or more models to

OpenAI’s red teaming innovations define new essentials for security leaders in the AI era

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More OpenAI has taken a more aggressive approach to red teaming than its AI competitors, demonstrating its security teams’ advanced capabilities in two areas: multi-step reinforcement and external red teaming. OpenAI recently released two papers that set a new competitive standard for improving the quality, reliability and safety of AI models in these two techniques and more. The first paper, “OpenAI’s Approach to External Red Teaming for AI Models and Systems,” reports that specialized teams outside the company have proven effective in uncovering vulnerabilities that might otherwise have made it into a released model because in-house testing techniques may have missed them. In the second paper, “Diverse and Effective Red Teaming with Auto-Generated Rewards and Multi-Step Reinforcement Learning,” OpenAI introduces an automated framework that relies on iterative reinforcement learning to generate a broad spectrum of novel, wide-ranging attacks. Going all-in on red teaming pays practical, competitive dividends It’s encouraging to see competitive intensity in red teaming growing among AI companies. When Anthropic released its AI red team guidelines in June of last year, it joined AI providers including Google, Microsoft, Nvidia, OpenAI, and even the U.S.’s National Institute of Standards and Technology (NIST), which all had released red teaming frameworks. Investing heavily in red teaming yields tangible benefits for security leaders in any organization. OpenAI’s paper on external red teaming provides a detailed analysis of how the company strives to create specialized external teams that include cybersecurity and subject matter experts. The goal is to see if knowledgeable external teams can defeat models’ security perimeters and find gaps in their security, biases and controls that prompt-based testing couldn’t find. What makes OpenAI’s recent papers noteworthy is how well they define using human-in-the-middle

Stay Ahead, Stay ONMINE