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.

Golar contracts CIMC Raffles for fourth FLNG

Golar LNG Ltd. last week executed an engineering, procurement, and construction (EPC) contract with Yantai CIMC Raffles Offshore Ltd. (CIMC Raffles) for its fourth floating LNG (FLNG) and second MKII design FLNG vessel with an annual liquefaction capacity of 3.5 million tonnes/year (tpy). Golar’s fourth FLNG is expected to be

Cerebras reimagines AI cluster design with switchless CS-4 architecture

Inside the Cerebras fabric, the trade-off is different. Direct Wafer Links are proprietary, so third-party accelerators cannot be added to the fabric. Scaling it requires additional Cerebras systems, Shah said. Faruqui said the operational divide may be more significant than the physical networking challenge. Existing orchestration environments are generally designed

IBM moves closer to fault-tolerant quantum computing at scale

“This is our first shared ultra cold environment that allows multiple chips within to be connected together, and it really provides enough space for all the high-density wiring that’s needed,” explained IBM Fellow Jerry Chow, CTO, quantum-centric supercomputing, during a media briefing. “So we’re really starting to architect the entire

Reports: Google partnering with AMD for next-gen hybrid TPU

For AMD, meanwhile, a role in Google’s next-generation TPU program would provide another indication that its CPU and chip-design technologies are becoming relevant to the rapidly expanding custom AI accelerator market. One analyst approved of the proposed alliance. “I think it makes good sense. AMD has more than proven that they’re

bp lets Shah Deniz compression automation contract

bp has let a contract to Emerson to deliver automation technologies for the Shah Deniz Compression project offshore Azerbaijan. Emerson will provide integrated control and safety systems aimed at enhancing production, safety, and reliability on the new offshore compression platform. The contract includes systems to provide process control, safety shutdown, fire and gas detection, and power management. Together, these systems deliver real-time visibility and remote control of critical operations, Emerson said. The $2.9 billion Shah Deniz Compression project, which includes an electrically powered, normally unattended offshore production platform, is a next stage development of the Caspian Sea Shah Deniz field. Designed to access low-pressure gas reserves and maximize overall recovery, the platform will be equipped with four 11 Mw compressors and serve as the central compression hub for gas from the Shah Deniz Alpha and Bravo platforms. The platform will operate remotely from bp’s onshore Sangachal terminal 55 km south of Baku. The project is expected to enable about 50 billion cu m of additional gas and about 25 million bbl of condensate production and export. Construction is scheduled to be completed in 2029, with first gas compression expected from the Shah Deniz Alpha platform in 2029 and from the Shah Deniz Bravo platform in 2030. The agreement follows a previous automation contract bp signed with Emerson for the Azeri Central East and Shah Deniz Stage 2 developments. bp is operator at Shah Deniz (29.99%) with partners Lukoil (19.99%), TPAO (19%), Cenub Qaz Dehlizi (16.02%), NICO (10%), and MVM (5%).

Federal court voids Texas GulfLink license over agency’s ‘serious procedural errors’

The ruling voids the license, halting all construction or progress. Sentinel Midstream declined comment on the ruling and would not answer questions about the status of construction. GulfLink, sited about 30 miles offshore Freeport, Tex., is designed to export up to 1 million b/d via Very Large Crude Carriers (VLCCs) to the government of Japan and Freeport Commodities. The project involves a 44-mile, 42-in. OD pipeline and was scheduled to begin operations around 2028. The estimated $2.1 billion investment was funded as part of a broader trade agreement between the US and Japan. The legal battle stems from a specific rule in the Deepwater Port Act of 1974 that dictates that the federal government can only permit one crude oil deepwater port, including any supporting infrastructure, within a single designated “application area.” Because the competing SPOT project’s pipeline route physically overlaps and intersects GulfLink’s lines, the plaintiff—Citizens for Clean Air & Clean Water in Brazoria County (Better Brazoria), represented by Earthjustice—successfully argued that MARAD violated the “one port” rule when issuing GulfLink’s license in February. The three-judge panel found that MARAD “improperly drew” the map designing the project’s official boundaries to exclude the pipelines and approved two overlapping projects in the same zone instead of only licensing one. The court wrote that the scope of the error made vacatur, not the less serious remand without vacatur, the appropriate remedy. Vacatur deems the license invalid and is used when the court finds “serious procedural errors” that cannot be easily explained or fixed with minor changes. Remand without vacatur sends the decision back to the agency for corrections but leaves the current license in place in the meantime. SPOT project status The $2.5-3-billion SPOT project, developed by Enterprise Products Partners in partnership with Enbridge Inc., also lies about 30 miles from Freeport. Designed to handle VLCCs,

IEA: Emergency reserve withdrawals slow

The International Energy Agency (IEA) member countries continued to release emergency oil stocks in July, but the pace of withdrawals slowed sharply as crude supply availability improved in parts of the Asia Pacific and the market faced increasing product tightness. IEA countries released 26 million bbl of emergency stocks in July, bringing cumulative releases to 300 million bbl since the agency announced a coordinated 400-million bbl action on Mar. 11. Government stock draws averaged 750,000 b/d in July, down from 1.5 million b/d in June and 2.5 million b/d in May. The slowdown was particularly pronounced among IEA members in Asia Oceania, which released 4 million bbl from emergency stocks in July, compared with 8 million bbl in June and 44 million bbl in May. The decline reflected improved crude oil supply availability in Japan and Korea. The US also reduced the pace of emergency stock releases. Withdrawals from the Strategic Petroleum Reserve totaled 17 million bbl in July, roughly half the volume released in June. More than 100 million bbl of the emergency stocks committed under the IEA’s 400-million bbl coordinated action has yet to reach the market. The timing of the remaining releases will depend on market developments and broader oil supply security considerations in coming months, according to the agency. Most of the remaining emergency stocks consist of crude oil, however, limiting their ability to ease increasingly tight oil product markets, IEA said. At the same time, global observed oil inventories fell sharply in July amid severely constrained shipping through the Strait of Hormuz. Stocks declined by 69 million bbl, equivalent to 2.2 million b/d, with oil on water accounting for more than 90% of the decline. Oil on water fell by 63 million bbl, or about 2 million b/d, reflecting higher arrivals and lower exports amid

Devon takes FID on 4.5-bcfd Permian pipeline

Devon Energy Corp. has taken final investment decision (FID) on the Solitude Pipeline System, a WhiteWater-led joint venture that will build two 48-in. OD natural gas pipelines connecting the Permian basin to Katy, Tex. Devon describes Solitude as the latest in a series of steps it has taken to integrate and consolidate the infrastructure supporting its Delaware basin position. Solitude is designed for a phased build-out of 2.25 bcfd entering service second-half 2029, followed by a similarly sized second phase in 2030 and the ability to expand further to meet shipper demand. Construction and in-service timing remain subject to customary regulatory approvals. Devon has secured firm transportation capacity and will hold a 25% equity interest in the joint venture, alongside WhiteWater (50%), MPLX (10%), Diamondback Energy (7.5%) and Western Midstream Partners (7.5%). Permian producers have long absorbed volatile and periodically negative pricing at the Waha hub, where takeaway capacity has repeatedly failed to keep pace with associated gas growth. Devon says the pipeline will move the majority of its Delaware gas out of Waha and into markets that will be tied to expanding LNG export and power generation. Devon has initiated the process of securing international LNG-linked pricing, including a 100-MMcfd agreement beginning in 2027 and an additional 150 MMcfd in 2028. “Solitude is not a standalone investment; it is the next step in an integrated model we have been building for years,” said Clay Gaspar, Devon’s president and chief executive officer. “We have taken the hardest constraints in the Delaware [b]asin: water, processing, compression, takeaway and power, and have de-risked the physical constraints turning each one into a source of value rather than a tax on our returns. The company’s integrated model continues to lower our cost of supply, driving free cash flow higher and deepening our peer-leading Delaware

JAPEX begins drilling second well for Tomakomai CCS Project

@import url(‘https://fonts.googleapis.com/css2?family=Inter:wght@100..900&display=swap’); .ebm-page__main h1, .ebm-page__main h2, .ebm-page__main h3, .ebm-page__main h4, .ebm-page__main h5, .ebm-page__main h6 { font-family: Inter; } body { line-height: 150%; letter-spacing: 0.025em; } button, .ebm-button-wrapper { font-family: Inter; } .label-style { text-transform: uppercase; color: var(–color-grey); font-weight: 600; font-size: 0.75rem; } .caption-style { font-size: 0.75rem; opacity: .6; } #onetrust-pc-sdk [id*=btn-handler], #onetrust-pc-sdk [class*=btn-handler] { background-color: #c19a06 !important; border-color: #c19a06 !important; } #onetrust-policy a, #onetrust-pc-sdk a, #ot-pc-content a { color: #c19a06 !important; } #onetrust-consent-sdk #onetrust-pc-sdk .ot-active-menu { border-color: #c19a06 !important; } #onetrust-consent-sdk #onetrust-accept-btn-handler, #onetrust-banner-sdk #onetrust-reject-all-handler, #onetrust-consent-sdk #onetrust-pc-btn-handler.cookie-setting-link { background-color: #c19a06 !important; border-color: #c19a06 !important; } #onetrust-consent-sdk .onetrust-pc-btn-handler { color: #c19a06 !important; border-color: #c19a06 !important; } Japan Petroleum Exploration (JAPEX) began drilling its second exploratory well for carbon dioxide (CO2) storage to advance the Tomakomai carbon capture and storage (CCS) project in the Tomakomai area of Hokkaido. Undertaken for drilling on Aug. 15, the well utilizes extended reach drilling from a drilling rig in Masago-cho, Tomakomai City, toward the seabed to confirm the existence of formations suitable for CO2 storage. Drilling of this second exploratory well follows start of exploratory drilling for the project in late 2025. Launched for initial study in 2023, the Tomakomai CCS project aims to separate and capture CO2 from Idemitsu Kosan Co. Ltd.’s Hokkaido refinery and Hokkaido Electric Power Co. Inc.’s Tomato Atsuma power station. The CO2 will be injected underground by JAPEX. While JAPEX said it intends to make a final investment decision (FID) on the Tomakomai CCS project based on combined results of both exploratory wells, the company did not specify a timeframe for when FID would occur. The partners previously said they intend to achieve a CO2-storage volume of 1.5-2 million tonnes/year in 2030.

APA efficiencies lead to lift in Permian production forecast

The leaders of APA Corp., Houston, have nudged up their 2026 forecast for oil production from the company’s US assets but kept their capital spending target level. APA’s production in the Permian basin, Egypt, and the North Sea totaled 410,000 boe/d during the second quarter, which was down from about 465,000 boe/d in the same period of last year as natural gas and international volumes, including by the company’s noncontrolling partner in Egypt, fell. Oil production in the Permian basin, where APA controls 159,000 net acres in the Delaware basin and 287,000 net acres in the Midland basin, was flat year over year at nearly 123,500, beating by 2% the estimates of chief executive officer John Christmann and his team. Speaking to analysts and investors on Aug. 6, Christmann said that APA’s drilling, completions, and field operations teams are growing more efficient and improving reliability. That, he added, is keeping the company on track for its target of saving $3.5 million per month in its operations by yearend. It also is letting APA executives tick up their US oil production outlook to 123,000 b/d from 122,000, echoing a similar move from 3 months ago. Projected US capex for the year is still $1.3 billion. “We’ve worked on adding durability and inventory life to the Permian, where we can run flat for more than 10 years,” Christmann said. “We’re obviously exceeding that with volumes and capital efficiency that we continue to have come through.” The APA team plans to keep production in the Permian basin and Egypt relatively flat as it prepares for first oil in 2028 from the GranMorgu project offshore Suriname, in which it is a 40% partner along with TotalEnergies (40%) and Staatsolie (20%). APA is putting to work $230 million in capital there this year, with some

Texas Tightens Oversight of Data Center Development

Texas has spent the past decade building one of the most data center-friendly policy environments in the United States. But the state’s political posture is tightening. The emerging message from Austin is that continued data center growth will face greater scrutiny over grid costs, water use, tax incentives and community impacts. What is interesting about this policy conversation is that the Texas Legislature is not in regular session. The 89th regular session ended June 2, 2025, and the 90th Legislature does not convene until January 12, 2027. What has occurred instead is a concentrated period of interim committee work, gubernatorial recommendations, implementation of Senate Bill 6, calls for a special session, and regulatory action by the Public Utility Commission of Texas and the Electric Reliability Council of Texas. Together, those efforts are creating the framework for a broader legislative debate in 2027 while already affecting projects seeking ERCOT interconnection, infrastructure costs and site-selection decisions. Abbott Sets Out a New Policy Framework The policy shift accelerated June 10, when Gov. Greg Abbott directed the PUCT to require data centers to fully fund the electric infrastructure needed to serve their operations and directed PUCT and ERCOT to identify additional actions available under existing authority. Separately, Abbott pledged to work with lawmakers in 2027 on legislation requiring data centers to add electric capacity, use water-efficient cooling systems, report electricity and water use, phase out outdated tax incentives and adopt additional protections for neighboring communities. The most consequential shift began June 10, when Gov. Greg Abbott sent state electricity regulators a sweeping list of data center policy priorities. Abbott called for future legislation requiring new facilities to add generation to the Texas grid, pay the full cost of their interconnection and related infrastructure, use closed-loop or similarly water-efficient cooling systems, and file annual reports

NVIDIA Pushes the AI Factory From Rack to Asset Class

Making Compute Underwritable Huang expanded the argument a day later in an NVIDIA blog describing AI factory compute as an emerging investable asset class. NVIDIA’s case begins with a definition. The company does not describe its compute platform simply as a GPU. It includes accelerated computing, networking, systems software, AI frameworks and the CUDA software ecosystem surrounding the hardware. That wider platform matters to the financing thesis because NVIDIA argues it increases the number of potential users for an installed AI system. An NVIDIA DSX AI factory could support language models, vision, speech, biological computing, robotics, physical AI and other workloads. The same infrastructure could potentially move among customers, clouds or operators as demand changes. In financial terms, NVIDIA is arguing for fungibility. That could become particularly important to lenders and infrastructure investors trying to determine what happens if an original customer disappears, a contract expires or the economics of a particular workload change. A GPU cluster tied economically to one speculative tenant is one thing. Compute that can be redeployed across a large global market of clouds, enterprises, AI developers and model providers is a different risk proposition. NVIDIA contends that this breadth of potential offtakers helps protect residual value. Whether institutional markets ultimately price that risk the way NVIDIA hopes remains to be seen. But the company is now explicitly trying to establish a financial framework around that premise. Challenging the Traditional Depreciation Curve NVIDIA’s second argument is that software can extend the economic life of installed hardware. CUDA is central to that case. The company maintains that successive software improvements can increase the performance and efficiency of systems that have already been deployed, allowing the same hardware to produce more useful work at lower cost over time. That does not eliminate hardware obsolescence. New GPU generations continue

The Next Data Center Constraint: Trust

When Facts Aren’t Enough Few places offer a more revealing test case than Loudoun County, Virginia. Data Center Alley has spent decades living with data center development at a scale most emerging markets will never approach. Rizer said Loudoun’s experience gives the county an unusually deep record with which to answer questions about environmental impacts, infrastructure and economic benefits. But those facts increasingly struggle to penetrate the broader public debate. Rizer said Loudoun today has more than 250 data centers, while the entire sector uses less than 10% of the county water system. He also pointed to improved air quality over the past decade and approximately $1.2 billion in tax revenue from the industry. Yet he acknowledged that simply producing another data point does little good when residents no longer trust the people presenting it. “I call it community concern whack-a-mole, because every time you address one thing, there are three others that pop up,” Rizer said. The problem, in his view, has become partly emotional rather than informational. “You can’t change how people think until you change how they feel,” he said. “And right now they feel angry, they feel confused, they are fearful, they are mistrustful, both of government and the big tech industry.” That distinction matters. The industry’s instinct has often been to counter criticism with facts: tax receipts, job numbers, water-use calculations, emissions data or explanations of how a particular cooling system works. Those facts remain important. But Rizer’s argument is that the industry must first rebuild enough credibility for communities to hear them. The Industry’s Unforced Errors Not all of the distrust has arrived from outside the industry. Rizer and Waitkunas were equally pointed about mistakes by developers and operators that have given opponents powerful examples to use against data center projects elsewhere. “The industry

Reports: Data Center Expansion Finds Its Contours

AI Density Is Arriving Unevenly Inside the data center, the AI transition remains equally uneven. Uptime’s 2026 survey found the average modal, or most common, rack density across respondents exceeding 11 kW for the first time, up from 9 kW in 2025. But that number requires context. A relatively small group of very high-density facilities pulls the average upward. Without those facilities, Uptime puts average modal rack density at 7.8 kW, only modestly higher than 7.5 kW in 2025. The industry therefore continues to operate two realities at once: a vast installed base running conventional rack densities and a rapidly emerging class of AI facilities pushing far beyond them. The latter is becoming more visible. Some 24% of Uptime respondents now report racks at 30 kW or higher, up from 19% last year. Much of the increase occurred above 50 kW, and some operators reported deployments exceeding 100 kW. Still, most surveyed facilities have no racks at 30 kW or above. AI inference is also moving up the density curve. For the first time in Uptime’s survey, generative AI inference matched AI training as a driver of respondents’ highest-density deployments, with 21% citing each workload. That matters because inference potentially pushes AI infrastructure requirements beyond a relatively concentrated population of model-training campuses and into a broader set of facilities and markets. Power Is Both Constraint and Risk No issue connects the three reports more consistently than power. It limits new site availability. It redirects development toward emerging markets. It shapes community debates. It affects density and cooling architecture. And once a facility is operating, it remains the largest source of outage risk. Uptime says 56% of operators who experienced an impactful outage identified power as the primary cause of their most recent incident. The institute cautions against treating the increase

DCF Poll: What Will Constrain Data Center Growth Next?

Matt Vincent is Editor in Chief of Data Center Frontier, where he leads editorial strategy and coverage focused on the infrastructure powering cloud computing, artificial intelligence, and the digital economy. A veteran B2B technology journalist with more than two decades of experience, Vincent specializes in the intersection of data centers, power, cooling, and emerging AI-era infrastructure. Since assuming the EIC role in 2023, he has helped guide Data Center Frontier’s coverage of the industry’s transition into the gigawatt-scale AI era, with a focus on hyperscale development, behind-the-meter power strategies, liquid cooling architectures, and the evolving energy demands of high-density compute, while working closely with the Digital Infrastructure Group at Endeavor Business Media to expand the brand’s analytical and multimedia footprint. Vincent also hosts The Data Center Frontier Show podcast, where he interviews industry leaders across hyperscale, colocation, utilities, and the data center supply chain to examine the technologies and business models reshaping digital infrastructure. Since its inception he serves as Head of Content for the Data Center Frontier Trends Summit. Before becoming Editor in Chief, he served in multiple senior editorial roles across Endeavor Business Media’s digital infrastructure portfolio, with coverage spanning data centers and hyperscale infrastructure, structured cabling and networking, telecom and datacom, IP physical security, and wireless and Pro AV markets. He began his career in 2005 within PennWell’s Advanced Technology Division and later held senior editorial positions supporting brands such as Cabling Installation & Maintenance, Lightwave Online, Broadband Technology Report, and Smart Buildings Technology. Vincent is a frequent moderator, interviewer, and keynote speaker at industry events including the HPC Forum, where he delivers forward-looking analysis on how AI and high-performance computing are reshaping digital infrastructure. He graduated with honors from Indiana University Bloomington with a B.A. in English Literature and Creative Writing and lives in southern New Hampshire with

Is your networking built for AI’s traffic patterns and data volumes?

As data centers evolve into AI factories, compute has shifted from a cost center to a revenue driver. “Compute is revenue,” said Jensen Huang, co-founder and CEO of NVIDIA. “Without compute, there is no way to generate tokens. Without tokens, there’s no way to generate revenue. So, in this new world of AI, compute equals revenue.” This reframe changes an organizations’ calculus. If compute is revenue, what do you optimize for? Here are 5 questions to consider: Are you measuring what actually drives AI factory revenue? Most AI factories are power-constrained, so tokens per watt dictate how much revenue you can generate and the cost per token impacts the AI factory profit margin. But neither of these metrics should be evaluated at a single operating point. Batch jobs, real-time chat, and agentic workloads demand different points on the throughput-latency curve. AI chips that perform well at only a few points will underserve the full range of workloads. Additional key operational metrics like time to first token (TTFT), mean time between interruptions (MTBI), and platform useful life are the bedrock of AI factory efficiency. They dictate how quickly an AI factory comes online to generate tokens, the reliability of its revenue streams, and its long-term ability to remain productive as AI workloads evolve. How does agentic AI change what your CPU needs to deliver? Data center CPUs have historically been optimized for parallel throughput, where more cores improve aggregate capacity. Agentic workloads run in loops and make different demands. The model reasons on the GPU, the CPU executes tool calls such as code compilation and data retrieval, and the result returns to the GPU so the model can reason again. Every step runs in sequence, gated by the one before it. Per-core performance and memory latency determine how fast each step

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