Bogdan Kulynych

Bogdan Kulynych, Juan Felipe Gomez, Georgios Kaissis et al.

Differential privacy (DP) is a widely used approach for mitigating privacy risks when training machine learning models on sensitive data. DP mechanisms add noise during training to limit the risk of information leakage. The scale of the added noise is critical, as it determines the trade-off between privacy and utility. The standard practice is to select the noise scale to satisfy a given privacy budget $\varepsilon$. This privacy budget is in turn interpreted in terms of operational attack risks, such as accuracy, sensitivity, and specificity of inference attacks aimed to recover information about the training data records. We show that first calibrating the noise scale to a privacy budget $\varepsilon$, and then translating ε to attack risk leads to overly conservative risk assessments and unnecessarily low utility. Instead, we propose methods to directly calibrate the noise scale to a desired attack risk level, bypassing the step of choosing $\varepsilon$. For a given notion of attack risk, our approach significantly decreases noise scale, leading to increased utility at the same level of privacy. We empirically demonstrate that calibrating noise to attack sensitivity/specificity, rather than $\varepsilon$, when training privacy-preserving ML models substantially improves model accuracy for the same risk level. Our work provides a principled and practical way to improve the utility of privacy-preserving ML without compromising on privacy. The code is available at https://github.com/Felipe-Gomez/riskcal

Bogdan Kulynych, Yao-Yuan Yang, Yaodong Yu et al. · berkeley, deepmind

Having similar behavior at training time and test time $-$ what we call a "What You See Is What You Get" (WYSIWYG) property $-$ is desirable in machine learning. Models trained with standard stochastic gradient descent (SGD), however, do not necessarily have this property, as their complex behaviors such as robustness or subgroup performance can differ drastically between training and test time. In contrast, we show that Differentially-Private (DP) training provably ensures the high-level WYSIWYG property, which we quantify using a notion of distributional generalization. Applying this connection, we introduce new conceptual tools for designing deep-learning methods by reducing generalization concerns to optimization ones: to mitigate unwanted behavior at test time, it is provably sufficient to mitigate this behavior on the training data. By applying this novel design principle, which bypasses "pathologies" of SGD, we construct simple algorithms that are competitive with SOTA in several distributional-robustness applications, significantly improve the privacy vs. disparate impact trade-off of DP-SGD, and mitigate robust overfitting in adversarial training. Finally, we also improve on theoretical bounds relating DP, stability, and distributional generalization.

Bogdan Kulynych, Hsiang Hsu, Carmela Troncoso et al.

Mechanisms used in privacy-preserving machine learning often aim to guarantee differential privacy (DP) during model training. Practical DP-ensuring training methods use randomization when fitting model parameters to privacy-sensitive data (e.g., adding Gaussian noise to clipped gradients). We demonstrate that such randomization incurs predictive multiplicity: for a given input example, the output predicted by equally-private models depends on the randomness used in training. Thus, for a given input, the predicted output can vary drastically if a model is re-trained, even if the same training dataset is used. The predictive-multiplicity cost of DP training has not been studied, and is currently neither audited for nor communicated to model designers and stakeholders. We derive a bound on the number of re-trainings required to estimate predictive multiplicity reliably. We analyze--both theoretically and through extensive experiments--the predictive-multiplicity cost of three DP-ensuring algorithms: output perturbation, objective perturbation, and DP-SGD. We demonstrate that the degree of predictive multiplicity rises as the level of privacy increases, and is unevenly distributed across individuals and demographic groups in the data. Because randomness used to ensure DP during training explains predictions for some examples, our results highlight a fundamental challenge to the justifiability of decisions supported by differentially private models in high-stakes settings. We conclude that practitioners should audit the predictive multiplicity of their DP-ensuring algorithms before deploying them in applications of individual-level consequence.

Klim Kireev, Bogdan Kulynych, Carmela Troncoso

Many safety-critical applications of machine learning, such as fraud or abuse detection, use data in tabular domains. Adversarial examples can be particularly damaging for these applications. Yet, existing works on adversarial robustness primarily focus on machine-learning models in image and text domains. We argue that, due to the differences between tabular data and images or text, existing threat models are not suitable for tabular domains. These models do not capture that the costs of an attack could be more significant than imperceptibility, or that the adversary could assign different values to the utility obtained from deploying different adversarial examples. We demonstrate that, due to these differences, the attack and defense methods used for images and text cannot be directly applied to tabular settings. We address these issues by proposing new cost and utility-aware threat models that are tailored to the adversarial capabilities and constraints of attackers targeting tabular domains. We introduce a framework that enables us to design attack and defense mechanisms that result in models protected against cost and utility-aware adversaries, for example, adversaries constrained by a certain financial budget. We show that our approach is effective on three datasets corresponding to applications for which adversarial examples can have economic and social implications.

Georgi Ganev, Meenatchi Sundaram Muthu Selva Annamalai, Bogdan Kulynych

State-of-the-art Differentially Private (DP) synthetic data generators such as MST and AIM are widely used, yet tightly auditing their privacy guarantees remains challenging. We introduce a Gaussian Differential Privacy (GDP)-based auditing framework that measures privacy via the full false-positive/false-negative tradeoff. Applied to MST and AIM under worst-case settings, our method provides the first tight audits in the strong-privacy regime. For $(ε,δ)=(1,10^{-2})$, we obtain $μ_{emp}\approx0.43$ vs. implied $μ=0.45$, showing a small theory-practice gap. Our code is publicly available: https://github.com/sassoftware/dpmm.

Avni Kothari, Bogdan Kulynych, Tsui-Wei Weng et al.

Machine learning models are often used to decide who receives a loan, a job interview, or a public benefit. Models in such settings use features without considering their actionability. As a result, they can assign predictions that are fixed $-$ meaning that individuals who are denied loans and interviews are, in fact, precluded from access to credit and employment. In this work, we introduce a procedure called recourse verification to test if a model assigns fixed predictions to its decision subjects. We propose a model-agnostic approach for recourse verification with reachable sets $-$ i.e., the set of all points that a person can reach through their actions in feature space. We develop methods to construct reachable sets for discrete feature spaces, which can certify the responsiveness of any model by simply querying its predictions. We conduct a comprehensive empirical study on the infeasibility of recourse on datasets from consumer finance. Our results highlight how models can inadvertently preclude access by assigning fixed predictions and underscore the need to account for actionability in model development.

Bogdan Kulynych, Theresa Stadler, Jean Louis Raisaro et al.

Recent advances in generative modelling have led many to see synthetic data as the go-to solution for a range of problems around data access, scarcity, and under-representation. In this paper, we study three prominent use cases: (1) Sharing synthetic data as a proxy for proprietary datasets to enable statistical analyses while protecting privacy, (2) Augmenting machine learning training sets with synthetic data to improve model performance, and (3) Augmenting datasets with synthetic data to reduce variance in statistical estimation. For each use case, we formalise the problem setting and study, through formal analysis and case studies, under which conditions synthetic data can achieve its intended objectives. We identify fundamental and practical limits that constrain when synthetic data can serve as an effective solution for a particular problem. Our analysis reveals that due to these limits many existing or envisioned use cases of synthetic data are a poor problem fit. Our formalisations and classification of synthetic data use cases enable decision makers to assess whether synthetic data is a suitable approach for their specific data availability problem.

Juan Felipe Gomez, Bogdan Kulynych, Georgios Kaissis et al.

Current practices for reporting the level of differential privacy (DP) protection for machine learning (ML) algorithms such as DP-SGD provide an incomplete and potentially misleading picture of the privacy guarantees. For instance, if only a single $(\varepsilon,δ)$ is known about a mechanism, standard analyses show that there exist highly accurate inference attacks against training data records, when, in fact, such accurate attacks might not exist. In this position paper, we argue that using non-asymptotic Gaussian Differential Privacy (GDP) as the primary means of communicating DP guarantees in ML avoids these potential downsides. Using two recent developments in the DP literature: (i) open-source numerical accountants capable of computing the privacy profile and $f$-DP curves of DP-SGD to arbitrary accuracy, and (ii) a decision-theoretic metric over DP representations, we show how to provide non-asymptotic bounds on GDP using numerical accountants, and show that GDP can capture the entire privacy profile of DP-SGD and related algorithms with virtually no error, as quantified by the metric. To support our claims, we investigate the privacy profiles of state-of-the-art DP large-scale image classification, and the TopDown algorithm for the U.S. Decennial Census, observing that GDP fits their profiles remarkably well in all cases. We conclude with a discussion on the strengths and weaknesses of this approach, and discuss which other privacy mechanisms could benefit from GDP.

Bogdan Kulynych, Mohammad Yaghini, Giovanni Cherubin et al.

A membership inference attack (MIA) against a machine-learning model enables an attacker to determine whether a given data record was part of the model's training data or not. In this paper, we provide an in-depth study of the phenomenon of disparate vulnerability against MIAs: unequal success rate of MIAs against different population subgroups. We first establish necessary and sufficient conditions for MIAs to be prevented, both on average and for population subgroups, using a notion of distributional generalization. Second, we derive connections of disparate vulnerability to algorithmic fairness and to differential privacy. We show that fairness can only prevent disparate vulnerability against limited classes of adversaries. Differential privacy bounds disparate vulnerability but can significantly reduce the accuracy of the model. We show that estimating disparate vulnerability to MIAs by naïvely applying existing attacks can lead to overestimation. We then establish which attacks are suitable for estimating disparate vulnerability, and provide a statistical framework for doing so reliably. We conduct experiments on synthetic and real-world data finding statistically significant evidence of disparate vulnerability in realistic settings. The code is available at https://github.com/spring-epfl/disparate-vulnerability

Bogdan Kulynych, Juan Felipe Gomez, Georgios Kaissis et al.

Differentially private (DP) mechanisms are difficult to interpret and calibrate because existing methods for mapping standard privacy parameters to concrete privacy risks -- re-identification, attribute inference, and data reconstruction -- are both overly pessimistic and inconsistent. In this work, we use the hypothesis-testing interpretation of DP ($f$-DP), and determine that bounds on attack success can take the same unified form across re-identification, attribute inference, and data reconstruction risks. Our unified bounds are (1) consistent across a multitude of attack settings, and (2) tunable, enabling practitioners to evaluate risk with respect to arbitrary, including worst-case, levels of baseline risk. Empirically, our results are tighter than prior methods using $\varepsilon$-DP, Rényi DP, and concentrated DP. As a result, calibrating noise using our bounds can reduce the required noise by 20% at the same risk level, which yields, e.g., an accuracy increase from 52% to 70% in a text classification task. Overall, this unifying perspective provides a principled framework for interpreting and calibrating the degree of protection in DP against specific levels of re-identification, attribute inference, or data reconstruction risk.

Theresa Stadler, Bogdan Kulynych, Michael C. Gastpar et al.

The promise of least-privilege learning -- to find feature representations that are useful for a learning task but prevent inference of any sensitive information unrelated to this task -- is highly appealing. However, so far this concept has only been stated informally. It thus remains an open question whether and how we can achieve this goal. In this work, we provide the first formalisation of the least-privilege principle for machine learning and characterise its feasibility. We prove that there is a fundamental trade-off between a representation's utility for a given task and its leakage beyond the intended task: it is not possible to learn representations that have high utility for the intended task but, at the same time prevent inference of any attribute other than the task label itself. This trade-off holds under realistic assumptions on the data distribution and regardless of the technique used to learn the feature mappings that produce these representations. We empirically validate this result for a wide range of learning techniques, model architectures, and datasets.

Giorgia Carra, Bogdan Kulynych, François Bastardot et al.

Although Large Language Models (LLMs) have shown promising performance in healthcare-related applications, their deployment in the medical domain poses unique challenges of ethical, regulatory, and technical nature. In this study, we employ a systematic participatory approach to investigate the needs and expectations regarding clinical applications of LLMs at Lausanne University Hospital, an academic medical center in Switzerland. Having identified potential LLM use-cases in collaboration with thirty stakeholders, including clinical staff across 11 departments as well nursing and patient representatives, we assess the current feasibility of these use-cases taking into account the regulatory frameworks, data protection regulation, bias, hallucinations, and deployment constraints. This study provides a framework for a participatory approach to identifying institutional needs with respect to introducing advanced technologies into healthcare practice, and a realistic analysis of the technology readiness level of LLMs for medical applications, highlighting the issues that would need to be overcome LLMs in healthcare to be ethical, and regulatory compliant.

Kendra Albert, Maggie Delano, Bogdan Kulynych et al.

Attacks from adversarial machine learning (ML) have the potential to be used "for good": they can be used to run counter to the existing power structures within ML, creating breathing space for those who would otherwise be the targets of surveillance and control. But most research on adversarial ML has not engaged in developing tools for resistance against ML systems. Why? In this paper, we review the broader impact statements that adversarial ML researchers wrote as part of their NeurIPS 2020 papers and assess the assumptions that authors have about the goals of their work. We also collect information about how authors view their work's impact more generally. We find that most adversarial ML researchers at NeurIPS hold two fundamental assumptions that will make it difficult for them to consider socially beneficial uses of attacks: (1) it is desirable to make systems robust, independent of context, and (2) attackers of systems are normatively bad and defenders of systems are normatively good. That is, despite their expressed and supposed neutrality, most adversarial ML researchers believe that the goal of their work is to secure systems, making it difficult to conceptualize and build tools for disrupting the status quo.

Agathe Balayn, Bogdan Kulynych, Seda Guerses

Researchers have identified datasets used for training computer vision (CV) models as an important source of hazardous outcomes, and continue to examine popular CV datasets to expose their harms. These works tend to treat datasets as objects, or focus on particular steps in data production pipelines. We argue here that we could further systematize our analysis of harms by examining CV data pipelines through a process-oriented lens that captures the creation, the evolution and use of these datasets. As a step towards cultivating a process-oriented lens, we embarked on an empirical study of CV data pipelines informed by the field of method engineering. We present here a preliminary result: a reference model of CV data pipelines. Besides exploring the questions that this endeavor raises, we discuss how the process lens could support researchers in discovering understudied issues, and could help practitioners in making their processes more transparent.

Rebekah Overdorf, Bogdan Kulynych, Ero Balsa et al.

In addition to their benefits, optimization systems can have negative economic, moral, social, and political effects on populations as well as their environments. Frameworks like fairness have been proposed to aid service providers in addressing subsequent bias and discrimination during data collection and algorithm design. However, recent reports of neglect, unresponsiveness, and malevolence cast doubt on whether service providers can effectively implement fairness solutions. These reports invite us to revisit assumptions made about the service providers in fairness solutions. Namely, that service providers have (i) the incentives or (ii) the means to mitigate optimization externalities. Moreover, the environmental impact of these systems suggests that we need (iii) novel frameworks that consider systems other than algorithmic decision-making and recommender systems, and (iv) solutions that go beyond removing related algorithmic biases. Going forward, we propose Protective Optimization Technologies that enable optimization subjects to defend against negative consequences of optimization systems.

Bogdan Kulynych, Jamie Hayes, Nikita Samarin et al.

Machine-learning models for security-critical applications such as bot, malware, or spam detection, operate in constrained discrete domains. These applications would benefit from having provable guarantees against adversarial examples. The existing literature on provable adversarial robustness of models, however, exclusively focuses on robustness to gradient-based attacks in domains such as images. These attacks model the adversarial cost, e.g., amount of distortion applied to an image, as a $p$-norm. We argue that this approach is not well-suited to model adversarial costs in constrained domains where not all examples are feasible. We introduce a graphical framework that (1) generalizes existing attacks in discrete domains, (2) can accommodate complex cost functions beyond $p$-norms, including financial cost incurred when attacking a classifier, and (3) efficiently produces valid adversarial examples with guarantees of minimal adversarial cost. These guarantees directly translate into a notion of adversarial robustness that takes into account domain constraints and the adversary's capabilities. We show how our framework can be used to evaluate security by crafting adversarial examples that evade a Twitter-bot detection classifier with provably minimal number of changes; and to build privacy defenses by crafting adversarial examples that evade a privacy-invasive website-fingerprinting classifier.

Bogdan Kulynych, Carmela Troncoso

Understanding the influence of features in machine learning is crucial to interpreting models and selecting the best features for classification. In this work we propose the use of principles from coalitional game theory to reason about importance of features. In particular, we propose the use of the Banzhaf power index as a measure of influence of features on the outcome of a classifier. We show that features having Banzhaf power index of zero can be losslessly pruned without damage to classifier accuracy. Computing the power indices does not require having access to data samples. However, if samples are available, the indices can be empirically estimated. We compute Banzhaf power indices for a neural network classifier on real-life data, and compare the results with gradient-based feature saliency, and coefficients of a logistic regression model with $L_1$ regularization.