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.

EPA to end environmental justice programs, monitoring tools

Dive Brief: The Trump administration announced Wednesday it will shut down all environmental justice offices and officially end other EJ-related initiatives, a move that will impact how waste and recycling industries measure and track their environmental impact on neighboring communities. The closures include the EPA’s Office of Environmental Justice and

Intel under Tan: What enterprise IT buyers need to know

Intel’s discrete GPU ambitions — especially in enterprise AI — have often appeared reactive rather than part of a clear strategic vision. The company entered the market late, facing Nvidia’s dominant CUDA ecosystem and AMD’s aggressive push into AI GPUs. “Tan’s background suggests he is unlikely to double down on

SUSE expands AI tools to control workloads, LLM usage

“And every few weeks we’ll continue to add to the library,” Puri says. SUSE also announced a partnership with Infosys today. The system integrator has the Topaz AI platform, which includes a set of services and solutions to help enterprises build and deploy AI applications. SUSE is also integrating the

D-Wave uses quantum to solve real-world problem

D-Wave published its results today, peer-reviewed in the journal Science. The classical supercomputer that D-Wave benchmarked against was the Frontier supercomputer at the Department of Energy’s Oak Ridge National Laboratory. It was, until recently, the most powerful supercomputer in the world but moved to second place in November. Two different

Crude Edges Higher After Seven Weeks of Declines

Oil snapped a seven-week losing streak as US equity markets rebounded and peace talks between Russia and Ukraine stalled, damping expectations that Moscow’s crude will return to the market soon. West Texas Intermediate rose almost 1% to settle above $67 a barrel, supported by a weaker dollar and an advance in US equities. Brent climbed to settle below $71. Russian President Vladimir Putin said Ukrainian troops in the Kursk region should lay down their arms, and Ukraine pushed back on the request, raising doubts about how soon a ceasefire could be achieved. US crude eked out a 0.2% gain for the week, barely skirting an eighth straight weekly decline that would have been its longest such losing streak since 2015. US President Donald Trump’s salvos against the country’s major trading partners have weighed on crude prices since mid-January, raising the prospect of sputtering economic growth and falling oil consumption. Long-term inflation expectations jumped by the most since 1993, painting a gloomy picture for future energy demand. US crude earlier rose as much as 1.4% after the White House imposed sanctions on Iran’s oil minister and on more companies and vessels used by the OPEC member, while also restricting payment options for Russian energy, before paring the gains. Still, the ceasefire negotiations unfolding between Russia and Ukraine, as well as macroeconomic risk, are holding traders’ attention for now, said Rebecca Babin, senior energy trader at CIBC Private Wealth Group. The sanctions developments are “all just words until they’re enforced, so the market is less reactive to the headlines recently,” Babin said. The potential return of Russian barrels comes amid projections the market already is headed for an oversupply. The IEA forecasts the global supply surplus is set to deepen as an escalating trade war pressures demand at the same time that

Qatar Supplies Syria With Natural Gas in Latest Post-Assad Boost

Qatar began supplying natural gas to Syria through Jordan, the latest boost to the war-torn country’s interim government following the fall of former president Bashar al-Assad. About 2 million cubic meters a day will be sent via the Arab Gas Pipeline, eventually contributing a total of 400 megawatts to the power grid, Syrian state-run news agency Sana said. The supplies were approved by Washington, Reuters reported earlier, without providing numbers. The contract signals further recognition for the government of Ahmed Al-Sharaa, who led the battle to overthrow Assad. It should help increase average power supply for Syrians to four hours a day, up from two, helping ease severe energy shortages. The UK removed the Syrian central bank and 23 other entities, mainly lenders and energy companies, from a list of sanctioned institutions earlier this month, following similar moves by several Western countries. Natural gas supplies through the Arab Gas Pipeline to Syria, and by extension to Lebanon, have been disrupted since 2011 due to the war and have been largely inactive since then. The exact mechanism by which Qatar will transport the gas to Syria and reactivate that section of the pipeline is unclear, as years of conflict have damaged vital energy infrastructure. Plus, the only LNG storage facility in Jordan, a vessel off the Red Sea port city of Aqaba, will be leased to Egypt for 10 years starting mid-2025. The power supply hinges on raising the production capacity of Syria’s Deir Ali power station, state-run Qatar News Agency said. This supply level is the “first phase” of a deal signed between Qatar Fund for Development and the Jordanian Ministry of Energy, in cooperation with the United Nations Development Program, which will oversee the “executive aspects of the project”. Syria’s interim government is seeking to replace oil imports from

Energy Bosses Shrug Off DeepSeek to Focus on Powering AI Boom

While tariffs and macroeconomic concerns weighed on the outlook for oil at a major energy conference in Houston this week, the mood around artificial intelligence and its sky-high power needs could scarcely be different. For a second year, energy executives at the CERAWeek by S&P Global gathering hailed the looming data center requirements for AI as both a huge challenge and a once-in-a-generation opportunity. “The only way we win the AI arms race with China is if we have electricity,” US Interior Secretary Doug Burgum said in his address. “They are moving at a speed that would suggest we are in a serious cyberwar with them.” The energy world appears to have shrugged off investor doubts that emerged over the AI-power narrative in January, when Chinese startup DeepSeek released a chat bot purported to use just a fraction of the electricity required by established US rivals. Despite that wobble, many forecasts for US power demand are still unprecedented — and come after more than two decades of stable consumption. Jenny Yang, head of power and renewables research at S&P, told conference delegates Thursday that US utilities’ estimates for additional power demand coming just from data centers by 2030 are equivalent to the entire Ercot power market in Texas. “We’re seeing load forecasts that, in my experience as a state regulator, are mind-boggling,” said Mark Christie, a former energy regulator in Virginia, the data-center capital of the US, and who now chairs the Federal Energy Regulatory Commission. The so-called hyperscalers continue to race ahead with their build-out of AI infrastructure. Google parent Alphabet Inc. reported last month it plans capital expenditures of $75 billion this year. The power demand related to that spending “is coming so fast and from so many different directions,” Alan Armstrong, chief executive officer of US pipeline operator Williams

The Emperor’s New Clothes: BP and Shell’s duck diplomacy

BP’s (LON:BP) undressing of its energy transition goals is the latest and most significant example of an oil supermajor reneging on its green investment pledges. It is easy to speculate that companies such as BP, and similarly Shell (LON:SHEL), have attempted to diversify into renewable energy too quickly. However, diversification in the energy transition could be the very thing that pulls the cart out of danger. This week, BP’s chief executive Murray Auchincloss defended the company’s decision to jettison renewable energy pledges and increase oil and gas production. In late February, he said the oil major had accelerated “too far, too fast” in the transition to renewable energy. “Our optimism for a fast transition was misplaced,” he said, after profits fell across its low-carbon and gas division, precipitating a sudden strategic about-face. The company, which has been under pressure from analysts and shareholders to reduce its low-carbon investments and double down on its core business of oil and gas, plans to cut investment in low-carbon projects by $5 billion (£4bn), Auchincloss said. © Image: BloombergLondon’s Old Oil Stocks Diverge | BP underperforms Shell on worries about green transition, payouts. “The challenge that faces BP and Equinor, and to varying degrees Shell and Equinor, is the marked underperformance of their shares relative to that of their US peers,” says Russ Mould, investment director at AJ Bell. “Whether this is down to the relatively greater emphasis they have placed upon investment in renewables to facilitate a move away from hydrocarbons or simply down to their stock market domicile (given how US equities continue to dominate across the board) is hard to divine, but the truth may well lie somewhere between. There is a sense that shareholders are becoming restless.” BP’s shares have shown a marked underperformance relative to global peers since former

Peterhead’s Acorn CCS key to unlocking future of Grangemouth

Grangemouth will need the Acorn Carbon Capture and Storage (CCS) development to go ahead to take full advantage of the upcoming £13 billion Project Willow plan. Colin Pritchard, sustainability and external relations director at Ineos, which runs the Grangemouth refinery Petroineos in a joint venture with PetroChina, said: “If you want to really go for all of the things that are within Willow and take them to the full extent, you will need a CO2 transportation and storage system. “In that case, the full extent of Willow needs Acorn.” Project Willow is the plan currently being developed by the UK and Scottish Governments to ameliorate the closure of the Scotland’s only oil refinery with the expected loss of 400 jobs. Due for release soon, Project Willow will lay out nine potential projects to overhaul the Grangemouth refinery in Scotland and create a long-term sustainable future for the site. A feasibility study exploring options for overhauling the Grangemouth refinery in Scotland is reportedly set to propose £3.8bn of investments in low-carbon alternatives for the site over ten years, with a best-case scenario could see the amount rise to almost £13bn. These options include recycling plastics, the production of biomethane, sustainable aviation fuel (SAF) and renewable diesel. In turn, these are hoped to avert the shutdown of Grangemouth, scheduled for the second quarter of this year, and preserve jobs at the facility. Speaking to Energy Voice on the side-lines of the DeCarbScotland event, Pritchard added: “There are some projects there are not dependent on Acorn, but there are some projects within Willow, like e-methanol, which are.” He added that the nine projects envisioned in Project Willow are an initial project set and could evolve, making CCS essential “if you want to get the full benefit of what we put in Willow”. Based in

EIA Reveals Latest Brent Oil Price Forecast for 2025 and 2026

The U.S. Energy Information Administration (EIA) has revealed its latest Brent spot price forecast for 2025 and 2026 in its March Short Term Energy Outlook (STEO), which was released this week. According to the STEO, the EIA now sees the Brent spot price averaging $74.22 per barrel this year and $68.47 per barrel next year. In its previous STEO, which was released in February, the EIA projected that the Brent spot price would average $74.50 per barrel in 2025 and $66.46 per barrel in 2026. The EIA outlined in its latest STEO that it sees the Brent spot price coming in at $74.89 per barrel in the first quarter of this year, $74.00 per barrel in the second quarter, $75.00 per barrel in the third quarter, $73.02 per barrel in the fourth quarter, $71.00 per barrel in the first quarter of 2026, $69.00 per barrel in the second quarter, $68.00 per barrel in the third quarter, and $66.00 per barrel in the fourth quarter. In its previous February STEO, the EIA forecast that the Brent spot price would average $77.13 per barrel in the first quarter of 2025, $75.00 per barrel in the second quarter, $74.00 per barrel in the third quarter, $72.00 per barrel in the fourth quarter, $68.97 per barrel in the first quarter of 2026, $67.33 per barrel in the second quarter, $65.68 per barrel in the third quarter, and $64.00 per barrel in the fourth quarter of next year. In its latest STEO, the EIA highlighted that the Brent crude oil spot price averaged $75 per barrel in February, which it pointed out was $4 per barrel lower than in January and $8 per barrel lower than at the same time last year. “Crude oil prices fell during February driven largely by economic growth concerns related

IBM laying foundation for mainframe as ultimate AI server

“It will truly change what customers are able to do with AI,” Stowell said. IBM’s mainframe processors The next generation of processors is expected to continue a long history of generation-to-generation improvements, IBM stated in a new white paper on AI and the mainframe. “They are projected to clock in at 5.5 GHz. and include ten 36 MB level 2 caches. They’ll feature built-in low-latency data processing for accelerated I/O as well as a completely redesigned cache and chip-interconnection infrastructure for more on-chip cache and compute capacity,” IBM wrote. Today’s mainframes also have extensions and accelerators that integrate with the core systems. These specialized add-ons are designed to enable the adoption of technologies such as Java, cloud and AI by accelerating computing paradigms that are essential for high-volume, low-latency transaction processing, IBM wrote. “The next crop of AI accelerators are expected to be significantly enhanced—with each accelerator designed to deliver 4 times more compute power, reaching 24 trillion operations per second (TOPS),” IBM wrote. “The I/O and cache improvements will enable even faster processing and analysis of large amounts of data and consolidation of workloads running across multiple servers, for savings in data center space and power costs. And the new accelerators will provide increased capacity to enable additional transaction clock time to perform enhanced in-transaction AI inferencing.” In addition, the next generation of the accelerator architecture is expected to be more efficient for AI tasks. “Unlike standard CPUs, the chip architecture will have a simpler layout, designed to send data directly from one compute engine, and use a range of lower- precision numeric formats. These enhancements are expected to make running AI models more energy efficient and far less memory intensive. As a result, mainframe users can leverage much more complex AI models and perform AI inferencing at a greater scale

VergeIO enhances VergeFabric network virtualization offering

VergeIO is not, however, using an off-the-shelf version of KVM. Rather, it is using what Crump referred to as a heavily modified KVM hypervisor base, with significant proprietary enhancements while still maintaining connections to the open-source community. VergeIO’s deployment profile is currently 70% on premises and about 30% via bare-metal service providers, with a particularly strong following among cloud service providers that host applications for their customers. The software requires direct hardware access due to its low-level integration with physical resources. “Since November of 2023, the normal number one customer we’re attracting right now is guys that have had a heart attack when they got their VMware renewal license,” Crump said. “The more of the stack you own, the better our story becomes.” A 2024 report from Data Center Intelligence Group (DCIG) identified VergeOS as one of the top 5 alternatives to VMware. “VergeIO starts by installing VergeOS on bare metal servers,” the report stated. “It then brings the servers’ hardware resources under its management, catalogs these resources, and makes them available to VMs. By directly accessing and managing the server’s hardware resources, it optimizes them in ways other hypervisors often cannot.” Advanced networking features in VergeFabric VergeFabric is the networking component within the VergeOS ecosystem, providing software-defined networking capabilities as an integrated service rather than as a separate virtual machine or application.

Podcast: On the Frontier of Modular Edge AI Data Centers with Flexnode’s Andrew Lindsey

The modular data center industry is undergoing a seismic shift in the age of AI, and few are as deeply embedded in this transformation as Andrew Lindsey, Co-Founder and CEO of Flexnode. In a recent episode of the Data Center Frontier Show podcast, Lindsey joined Editor-in-Chief Matt Vincent and Senior Editor David Chernicoff to discuss the evolution of modular data centers, the growing demand for high-density liquid-cooled solutions, and the industry factors driving this momentum. A Background Rooted in Innovation Lindsey’s career has been defined by the intersection of technology and the built environment. Prior to launching Flexnode, he worked at Alpha Corporation, a top 100 engineering and construction management firm founded by his father in 1979. His early career involved spearheading technology adoption within the firm, with a focus on high-security infrastructure for both government and private clients. Recognizing a massive opportunity in the data center space, Lindsey saw a need for an innovative approach to infrastructure deployment. “The construction industry is relatively uninnovative,” he explained, citing a McKinsey study that ranked construction as the second least-digitized industry—just above fishing and wildlife, which remains deliberately undigitized. Given the billions of square feet of data center infrastructure required in a relatively short timeframe, Lindsey set out to streamline and modernize the process. Founded four years ago, Flexnode delivers modular data centers with a fully integrated approach, handling everything from site selection to design, engineering, manufacturing, deployment, operations, and even end-of-life decommissioning. Their core mission is to provide an “easy button” for high-density computing solutions, including cloud and dedicated GPU infrastructure, allowing faster and more efficient deployment of modular data centers. The Rising Momentum for Modular Data Centers As Vincent noted, Data Center Frontier has closely tracked the increasing traction of modular infrastructure. Lindsey has been at the forefront of this

Last Energy to Deploy 30 Microreactors in Texas for Data Centers

As the demand for data center power surges in Texas, nuclear startup Last Energy has now announced plans to build 30 microreactors in the state’s Haskell County near the Dallas-Fort Worth Metroplex. The reactors will serve a growing customer base of data center operators in the region looking for reliable, carbon-free energy. The plan marks Last Energy’s largest project to date and a significant step in advancing modular nuclear power as a viable solution for high-density computing infrastructure. Meeting the Looming Power Demands of Texas Data Centers Texas is already home to over 340 data centers, with significant expansion underway. Google is increasing its data center footprint in Dallas, while OpenAI’s Stargate has announced plans for a new facility in Abilene, just an hour south of Last Energy’s planned site. The company notes the Dallas-Fort Worth metro area alone is projected to require an additional 43 gigawatts of power in the coming years, far surpassing current grid capacity. To help remediate, Last Energy has secured a 200+ acre site in Haskell County, approximately three and a half hours west of Dallas. The company has also filed for a grid connection with ERCOT, with plans to deliver power via a mix of private wire and grid transmission. Additionally, Last Energy has begun pre-application engagement with the U.S. Nuclear Regulatory Commission (NRC) for an Early Site Permit, a key step in securing regulatory approval. According to Last Energy CEO Bret Kugelmass, the company’s modular approach is designed to bring nuclear energy online faster than traditional projects. “Nuclear power is the most effective way to meet Texas’ growing energy demand, but it needs to be deployed faster and at scale,” Kugelmass said. “Our microreactors are designed to be plug-and-play, enabling data center operators to bypass the constraints of an overloaded grid.” Scaling Nuclear for

Data Center Jobs: Engineering and Technician Jobs Available in Major Markets

Each month Data Center Frontier, in partnership with Pkaza, posts some of the hottest data center career opportunities in the market. Here’s a look at some of the latest data center jobs posted on the Data Center Frontier jobs board, powered by Pkaza Critical Facilities Recruiting. Data Center Facility Engineer (Night Shift Available) Ashburn, VAThis position is also available in: Tacoma, WA (Nights), Days/Nights: Needham, MA and New York City, NY. This opportunity is working directly with a leading mission-critical data center developer / wholesaler / colo provider. This firm provides data center solutions custom-fit to the requirements of their client’s mission-critical operational facilities. They provide reliability of mission-critical facilities for many of the world’s largest organizations facilities supporting enterprise clients and hyperscale companies. This opportunity provides a career-growth minded role with exciting projects with leading-edge technology and innovation as well as competitive salaries and benefits. Electrical Commissioning Engineer New Albany, OHThis traveling position is also available in: Somerset, NJ; Boydton, VA; Richmond, VA; Ashburn, VA; Charlotte, NC; Atlanta, GA; Hampton, GA; Fayetteville, GA; Des Moines, IA; San Jose, CA; Portland, OR; St Louis, MO; Phoenix, AZ; Dallas, TX; Chicago, IL; or Toronto, ON. *** ALSO looking for a LEAD EE and ME CxA agents.*** Our client is an engineering design and commissioning company that has a national footprint and specializes in MEP critical facilities design. They provide design, commissioning, consulting and management expertise in the critical facilities space. They have a mindset to provide reliability, energy efficiency, sustainable design and LEED expertise when providing these consulting services for enterprise, colocation and hyperscale companies. This career-growth minded opportunity offers exciting projects with leading-edge technology and innovation as well as competitive salaries and benefits. Switchgear Field Service Technician – Critical Facilities Nationwide TravelThis position is also available in: Charlotte, NC; Atlanta, GA; Dallas,

Amid Shifting Regional Data Center Policies, Iron Mountain and DC Blox Both Expand in Virginia’s Henrico County

The dynamic landscape of data center developments in Maryland and Virginia exemplify the intricate balance between fostering technological growth and addressing community and environmental concerns. Data center developers in this region find themselves both in the crosshairs of groups worried about the environment and other groups looking to drive economic growth. In some cases, the groups are different components of the same organizations, such as local governments. For data center development, meeting the needs of these competing interests often means walking a none-too-stable tightrope. Rapid Government Action Encourages Growth In May 2024, Maryland demonstrated its commitment to attracting data center investments by enacting the Critical Infrastructure Streamlining Act. This legislation provides a clear framework for the use of emergency backup power generation, addressing previous regulatory challenges that a few months earlier had hindered projects like Aligned Data Centers’ proposed 264-megawatt campus in Frederick County, causing Aligned to pull out of the project. However, just days after the Act was signed by the governor, Aligned reiterated its plans to move forward with development in Maryland. With the Quantum Loop and the related data center development making Frederick County a focal point for a balanced approach, the industry is paying careful attention to the pace of development and the relations between developers, communities and the government. In September of 2024, Frederick County Executive Jessica Fitzwater revealed draft legislation that would potentially restrict where in the county data centers could be built. The legislation was based on information found in the Frederick County Data Centers Workgroup’s final report. Those bills would update existing regulations and create a floating zone for Critical Digital Infrastructure and place specific requirements on siting data centers. Statewide, a cautious approach to environmental and community impacts statewide has been deemed important. In January 2025, legislators introduced SB116, a bill

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