Somesh Jha

Clark Barrett, Brad Boyd, Elie Burzstein et al. · berkeley

Every major technical invention resurfaces the dual-use dilemma -- the new technology has the potential to be used for good as well as for harm. Generative AI (GenAI) techniques, such as large language models (LLMs) and diffusion models, have shown remarkable capabilities (e.g., in-context learning, code-completion, and text-to-image generation and editing). However, GenAI can be used just as well by attackers to generate new attacks and increase the velocity and efficacy of existing attacks. This paper reports the findings of a workshop held at Google (co-organized by Stanford University and the University of Wisconsin-Madison) on the dual-use dilemma posed by GenAI. This paper is not meant to be comprehensive, but is rather an attempt to synthesize some of the interesting findings from the workshop. We discuss short-term and long-term goals for the community on this topic. We hope this paper provides both a launching point for a discussion on this important topic as well as interesting problems that the research community can work to address.

Samuel Maddock, Graham Cormode, Tianhao Wang et al. · oxford

There is great demand for scalable, secure, and efficient privacy-preserving machine learning models that can be trained over distributed data. While deep learning models typically achieve the best results in a centralized non-secure setting, different models can excel when privacy and communication constraints are imposed. Instead, tree-based approaches such as XGBoost have attracted much attention for their high performance and ease of use; in particular, they often achieve state-of-the-art results on tabular data. Consequently, several recent works have focused on translating Gradient Boosted Decision Tree (GBDT) models like XGBoost into federated settings, via cryptographic mechanisms such as Homomorphic Encryption (HE) and Secure Multi-Party Computation (MPC). However, these do not always provide formal privacy guarantees, or consider the full range of hyperparameters and implementation settings. In this work, we implement the GBDT model under Differential Privacy (DP). We propose a general framework that captures and extends existing approaches for differentially private decision trees. Our framework of methods is tailored to the federated setting, and we show that with a careful choice of techniques it is possible to achieve very high utility while maintaining strong levels of privacy.

Saeed Mahloujifar, Alexandre Sablayrolles, Graham Cormode et al. · oxford

Given a trained model and a data sample, membership-inference (MI) attacks predict whether the sample was in the model's training set. A common countermeasure against MI attacks is to utilize differential privacy (DP) during model training to mask the presence of individual examples. While this use of DP is a principled approach to limit the efficacy of MI attacks, there is a gap between the bounds provided by DP and the empirical performance of MI attacks. In this paper, we derive bounds for the \textit{advantage} of an adversary mounting a MI attack, and demonstrate tightness for the widely-used Gaussian mechanism. We further show bounds on the \textit{confidence} of MI attacks. Our bounds are much stronger than those obtained by DP analysis. For example, analyzing a setting of DP-SGD with $ε=4$ would obtain an upper bound on the advantage of $\approx0.36$ based on our analyses, while getting bound of $\approx 0.97$ using the analysis of previous work that convert $ε$ to membership inference bounds. Finally, using our analysis, we provide MI metrics for models trained on CIFAR10 dataset. To the best of our knowledge, our analysis provides the state-of-the-art membership inference bounds for the privacy.

Debopam Sanyal, Jui-Tse Hung, Manav Agrawal et al.

Model-serving systems have become increasingly popular, especially in real-time web applications. In such systems, users send queries to the server and specify the desired performance metrics (e.g., desired accuracy, latency). The server maintains a set of models (model zoo) in the back-end and serves the queries based on the specified metrics. This paper examines the security, specifically robustness against model extraction attacks, of such systems. Existing black-box attacks assume a single model can be repeatedly selected for serving inference requests. Modern inference serving systems break this assumption. Thus, they cannot be directly applied to extract a victim model, as models are hidden behind a layer of abstraction exposed by the serving system. An attacker can no longer identify which model she is interacting with. To this end, we first propose a query-efficient fingerprinting algorithm to enable the attacker to trigger any desired model consistently. We show that by using our fingerprinting algorithm, model extraction can have fidelity and accuracy scores within $1\%$ of the scores obtained when attacking a single, explicitly specified model, as well as up to $14.6\%$ gain in accuracy and up to $7.7\%$ gain in fidelity compared to the naive attack. Second, we counter the proposed attack with a noise-based defense mechanism that thwarts fingerprinting by adding noise to the specified performance metrics. The proposed defense strategy reduces the attack's accuracy and fidelity by up to $9.8\%$ and $4.8\%$, respectively (on medium-sized model extraction). Third, we show that the proposed defense induces a fundamental trade-off between the level of protection and system goodput, achieving configurable and significant victim model extraction protection while maintaining acceptable goodput ($>80\%$). We implement the proposed defense in a real system with plans to open source.

Adam Dziedzic, Christopher A Choquette-Choo, Natalie Dullerud et al. · deepmind, utoronto

Private multi-winner voting is the task of revealing $k$-hot binary vectors satisfying a bounded differential privacy (DP) guarantee. This task has been understudied in machine learning literature despite its prevalence in many domains such as healthcare. We propose three new DP multi-winner mechanisms: Binary, $τ$, and Powerset voting. Binary voting operates independently per label through composition. $τ$ voting bounds votes optimally in their $\ell_2$ norm for tight data-independent guarantees. Powerset voting operates over the entire binary vector by viewing the possible outcomes as a power set. Our theoretical and empirical analysis shows that Binary voting can be a competitive mechanism on many tasks unless there are strong correlations between labels, in which case Powerset voting outperforms it. We use our mechanisms to enable privacy-preserving multi-label learning in the central setting by extending the canonical single-label technique: PATE. We find that our techniques outperform current state-of-the-art approaches on large, real-world healthcare data and standard multi-label benchmarks. We further enable multi-label confidential and private collaborative (CaPC) learning and show that model performance can be significantly improved in the multi-site setting.

Olive Franzese, Adam Dziedzic, Christopher A. Choquette-Choo et al. · deepmind

Collaborative machine learning (ML) is widely used to enable institutions to learn better models from distributed data. While collaborative approaches to learning intuitively protect user data, they remain vulnerable to either the server, the clients, or both, deviating from the protocol. Indeed, because the protocol is asymmetric, a malicious server can abuse its power to reconstruct client data points. Conversely, malicious clients can corrupt learning with malicious updates. Thus, both clients and servers require a guarantee when the other cannot be trusted to fully cooperate. In this work, we propose a peer-to-peer (P2P) learning scheme that is secure against malicious servers and robust to malicious clients. Our core contribution is a generic framework that transforms any (compatible) algorithm for robust aggregation of model updates to the setting where servers and clients can act maliciously. Finally, we demonstrate the computational efficiency of our approach even with 1-million parameter models trained by 100s of peers on standard datasets.

Zhenmei Shi, Jiefeng Chen, Kunyang Li et al.

Pre-training representations (a.k.a. foundation models) has recently become a prevalent learning paradigm, where one first pre-trains a representation using large-scale unlabeled data, and then learns simple predictors on top of the representation using small labeled data from the downstream tasks. There are two key desiderata for the representation: label efficiency (the ability to learn an accurate classifier on top of the representation with a small amount of labeled data) and universality (usefulness across a wide range of downstream tasks). In this paper, we focus on one of the most popular instantiations of this paradigm: contrastive learning with linear probing, i.e., learning a linear predictor on the representation pre-trained by contrastive learning. We show that there exists a trade-off between the two desiderata so that one may not be able to achieve both simultaneously. Specifically, we provide analysis using a theoretical data model and show that, while more diverse pre-training data result in more diverse features for different tasks (improving universality), it puts less emphasis on task-specific features, giving rise to larger sample complexity for down-stream supervised tasks, and thus worse prediction performance. Guided by this analysis, we propose a contrastive regularization method to improve the trade-off. We validate our analysis and method empirically with systematic experiments using real-world datasets and foundation models.

Jihye Choi, Jayaram Raghuram, Ryan Feng et al.

Out-of-distribution (OOD) detection plays a crucial role in ensuring the safe deployment of deep neural network (DNN) classifiers. While a myriad of methods have focused on improving the performance of OOD detectors, a critical gap remains in interpreting their decisions. We help bridge this gap by providing explanations for OOD detectors based on learned high-level concepts. We first propose two new metrics for assessing the effectiveness of a particular set of concepts for explaining OOD detectors: 1) detection completeness, which quantifies the sufficiency of concepts for explaining an OOD-detector's decisions, and 2) concept separability, which captures the distributional separation between in-distribution and OOD data in the concept space. Based on these metrics, we propose an unsupervised framework for learning a set of concepts that satisfy the desired properties of high detection completeness and concept separability, and demonstrate its effectiveness in providing concept-based explanations for diverse off-the-shelf OOD detectors. We also show how to identify prominent concepts contributing to the detection results, and provide further reasoning about their decisions.

Zi Wang, Gautam Prakriya, Somesh Jha

Fast and precise Lipschitz constant estimation of neural networks is an important task for deep learning. Researchers have recently found an intrinsic trade-off between the accuracy and smoothness of neural networks, so training a network with a loose Lipschitz constant estimation imposes a strong regularization and can hurt the model accuracy significantly. In this work, we provide a unified theoretical framework, a quantitative geometric approach, to address the Lipschitz constant estimation. By adopting this framework, we can immediately obtain several theoretical results, including the computational hardness of Lipschitz constant estimation and its approximability. Furthermore, the quantitative geometric perspective can also provide some insights into recent empirical observations that techniques for one norm do not usually transfer to another one. We also implement the algorithms induced from this quantitative geometric approach in a tool GeoLIP. These algorithms are based on semidefinite programming (SDP). Our empirical evaluation demonstrates that GeoLIP is more scalable and precise than existing tools on Lipschitz constant estimation for $\ell_\infty$-perturbations. Furthermore, we also show its intricate relations with other recent SDP-based techniques, both theoretically and empirically. We believe that this unified quantitative geometric perspective can bring new insights and theoretical tools to the investigation of neural-network smoothness and robustness.

Jaiden Fairoze, Sanjam Garg, Somesh Jha et al.

We present a publicly-detectable watermarking scheme for LMs: the detection algorithm contains no secret information, and it is executable by anyone. We embed a publicly-verifiable cryptographic signature into LM output using rejection sampling and prove that this produces unforgeable and distortion-free (i.e., undetectable without access to the public key) text output. We make use of error-correction to overcome periods of low entropy, a barrier for all prior watermarking schemes. We implement our scheme and find that our formal claims are met in practice.

Ryan Feng, Ashish Hooda, Neal Mangaokar et al.

Recent work has proposed stateful defense models (SDMs) as a compelling strategy to defend against a black-box attacker who only has query access to the model, as is common for online machine learning platforms. Such stateful defenses aim to defend against black-box attacks by tracking the query history and detecting and rejecting queries that are "similar" and thus preventing black-box attacks from finding useful gradients and making progress towards finding adversarial attacks within a reasonable query budget. Recent SDMs (e.g., Blacklight and PIHA) have shown remarkable success in defending against state-of-the-art black-box attacks. In this paper, we show that SDMs are highly vulnerable to a new class of adaptive black-box attacks. We propose a novel adaptive black-box attack strategy called Oracle-guided Adaptive Rejection Sampling (OARS) that involves two stages: (1) use initial query patterns to infer key properties about an SDM's defense; and, (2) leverage those extracted properties to design subsequent query patterns to evade the SDM's defense while making progress towards finding adversarial inputs. OARS is broadly applicable as an enhancement to existing black-box attacks - we show how to apply the strategy to enhance six common black-box attacks to be more effective against current class of SDMs. For example, OARS-enhanced versions of black-box attacks improved attack success rate against recent stateful defenses from almost 0% to to almost 100% for multiple datasets within reasonable query budgets.

Ashish Hooda, Rishabh Khandelwal, Prasad Chalasani et al.

Privacy policies are crucial in the online ecosystem, defining how services handle user data and adhere to regulations such as GDPR and CCPA. However, their complexity and frequent updates often make them difficult for stakeholders to understand and analyze. Current automated analysis methods, which utilize natural language processing, have limitations. They typically focus on individual tasks and fail to capture the full context of the policies. We propose PolicyLR, a new paradigm that offers a comprehensive machine-readable representation of privacy policies, serving as an all-in-one solution for multiple downstream tasks. PolicyLR converts privacy policies into a machine-readable format using valuations of atomic formulae, allowing for formal definitions of tasks like compliance and consistency. We have developed a compiler that transforms unstructured policy text into this format using off-the-shelf Large Language Models (LLMs). This compiler breaks down the transformation task into a two-stage translation and entailment procedure. This procedure considers the full context of the privacy policy to infer a complex formula, where each formula consists of simpler atomic formulae. The advantage of this model is that PolicyLR is interpretable by design and grounded in segments of the privacy policy. We evaluated the compiler using ToS;DR, a community-annotated privacy policy entailment dataset. Utilizing open-source LLMs, our compiler achieves precision and recall values of 0.91 and 0.88, respectively. Finally, we demonstrate the utility of PolicyLR in three privacy tasks: Policy Compliance, Inconsistency Detection, and Privacy Comparison Shopping.

Jiefeng Chen, Jinsung Yoon, Sayna Ebrahimi et al.

Large language models (LLMs) have recently shown great advances in a variety of tasks, including natural language understanding and generation. However, their use in high-stakes decision-making scenarios is still limited due to the potential for errors. Selective prediction is a technique that can be used to improve the reliability of the LLMs by allowing them to abstain from making predictions when they are unsure of the answer. In this work, we propose a novel framework for adaptation with self-evaluation to improve the selective prediction performance of LLMs. Our framework is based on the idea of using parameter-efficient tuning to adapt the LLM to the specific task at hand while improving its ability to perform self-evaluation. We evaluate our method on a variety of question-answering (QA) datasets and show that it outperforms state-of-the-art selective prediction methods. For example, on the CoQA benchmark, our method improves the AUACC from 91.23% to 92.63% and improves the AUROC from 74.61% to 80.25%.

Jiefeng Chen, Jinsung Yoon, Sayna Ebrahimi et al.

Selective prediction aims to learn a reliable model that abstains from making predictions when uncertain. These predictions can then be deferred to humans for further evaluation. As an everlasting challenge for machine learning, in many real-world scenarios, the distribution of test data is different from the training data. This results in more inaccurate predictions, and often increased dependence on humans, which can be difficult and expensive. Active learning aims to lower the overall labeling effort, and hence human dependence, by querying the most informative examples. Selective prediction and active learning have been approached from different angles, with the connection between them missing. In this work, we introduce a new learning paradigm, active selective prediction, which aims to query more informative samples from the shifted target domain while increasing accuracy and coverage. For this new paradigm, we propose a simple yet effective approach, ASPEST, that utilizes ensembles of model snapshots with self-training with their aggregated outputs as pseudo labels. Extensive experiments on numerous image, text and structured datasets, which suffer from domain shifts, demonstrate that ASPEST can significantly outperform prior work on selective prediction and active learning (e.g. on the MNIST$\to$SVHN benchmark with the labeling budget of 100, ASPEST improves the AUACC metric from 79.36% to 88.84%) and achieves more optimal utilization of humans in the loop.

Ashish Hooda, Neal Mangaokar, Ryan Feng et al.

Adversarial examples threaten the integrity of machine learning systems with alarming success rates even under constrained black-box conditions. Stateful defenses have emerged as an effective countermeasure, detecting potential attacks by maintaining a buffer of recent queries and detecting new queries that are too similar. However, these defenses fundamentally pose a trade-off between attack detection and false positive rates, and this trade-off is typically optimized by hand-picking feature extractors and similarity thresholds that empirically work well. There is little current understanding as to the formal limits of this trade-off and the exact properties of the feature extractors/underlying problem domain that influence it. This work aims to address this gap by offering a theoretical characterization of the trade-off between detection and false positive rates for stateful defenses. We provide upper bounds for detection rates of a general class of feature extractors and analyze the impact of this trade-off on the convergence of black-box attacks. We then support our theoretical findings with empirical evaluations across multiple datasets and stateful defenses.

Sanjam Garg, Somesh Jha, Saeed Mahloujifar et al.

Overparameterized models with millions of parameters have been hugely successful. In this work, we ask: can the need for large models be, at least in part, due to the \emph{computational} limitations of the learner? Additionally, we ask, is this situation exacerbated for \emph{robust} learning? We show that this indeed could be the case. We show learning tasks for which computationally bounded learners need \emph{significantly more} model parameters than what information-theoretic learners need. Furthermore, we show that even more model parameters could be necessary for robust learning. In particular, for computationally bounded learners, we extend the recent result of Bubeck and Sellke [NeurIPS'2021] which shows that robust models might need more parameters, to the computational regime and show that bounded learners could provably need an even larger number of parameters. Then, we address the following related question: can we hope to remedy the situation for robust computationally bounded learning by restricting \emph{adversaries} to also be computationally bounded for sake of obtaining models with fewer parameters? Here again, we show that this could be possible. Specifically, building on the work of Garg, Jha, Mahloujifar, and Mahmoody [ALT'2020], we demonstrate a learning task that can be learned efficiently and robustly against a computationally bounded attacker, while to be robust against an information-theoretic attacker requires the learner to utilize significantly more parameters.

Zi Wang, Somesh Jha, Krishnamurthy et al.

The neural network has become an integral part of modern software systems. However, they still suffer from various problems, in particular, vulnerability to adversarial attacks. In this work, we present a novel program reasoning framework for neural-network verification, which we refer to as symbolic reasoning. The key components of our framework are the use of the symbolic domain and the quadratic relation. The symbolic domain has very flexible semantics, and the quadratic relation is quite expressive. They allow us to encode many verification problems for neural networks as quadratic programs. Our scheme then relaxes the quadratic programs to semidefinite programs, which can be efficiently solved. This framework allows us to verify various neural-network properties under different scenarios, especially those that appear challenging for non-symbolic domains. Moreover, it introduces new representations and perspectives for the verification tasks. We believe that our framework can bring new theoretical insights and practical tools to verification problems for neural networks.

Mingtian Tan, Tianhao Wang, Somesh Jha

Recently, diffusion models (DMs) have become the state-of-the-art method for image synthesis. Editing models based on DMs, known for their high fidelity and precision, have inadvertently introduced new challenges related to image copyright infringement and malicious editing. Our work is the first to formalize and address this issue. After assessing and attempting to enhance traditional image watermarking techniques, we recognize their limitations in this emerging context. In response, we develop a novel technique, RIW (Robust Invisible Watermarking), to embed invisible watermarks leveraging adversarial example techniques. Our technique ensures a high extraction accuracy of $96\%$ for the invisible watermark after editing, compared to the $0\%$ offered by conventional methods. We provide access to our code at https://github.com/BennyTMT/RIW.

Matt Fredrikson, Kaiji Lu, Saranya Vijayakumar et al.

Recent techniques that integrate \emph{solver layers} into Deep Neural Networks (DNNs) have shown promise in bridging a long-standing gap between inductive learning and symbolic reasoning techniques. In this paper we present a set of techniques for integrating \emph{Satisfiability Modulo Theories} (SMT) solvers into the forward and backward passes of a deep network layer, called SMTLayer. Using this approach, one can encode rich domain knowledge into the network in the form of mathematical formulas. In the forward pass, the solver uses symbols produced by prior layers, along with these formulas, to construct inferences; in the backward pass, the solver informs updates to the network, driving it towards representations that are compatible with the solver's theory. Notably, the solver need not be differentiable. We implement \layername as a Pytorch module, and our empirical results show that it leads to models that \emph{1)} require fewer training samples than conventional models, \emph{2)} that are robust to certain types of covariate shift, and \emph{3)} that ultimately learn representations that are consistent with symbolic knowledge, and thus naturally interpretable.

Jihye Choi, Shruti Tople, Varun Chandrasekaran et al.

Membership Inference Attacks (MIAs) aim to identify specific data samples within the private training dataset of machine learning models, leading to serious privacy violations and other sophisticated threats. Many practical black-box MIAs require query access to the data distribution (the same distribution where the private data is drawn) to train shadow models. By doing so, the adversary obtains models trained "with" or "without" samples drawn from the distribution, and analyzes the characteristics of the samples under consideration. The adversary is often required to train more than hundreds of shadow models to extract the signals needed for MIAs; this becomes the computational overhead of MIAs. In this paper, we propose that by strategically choosing the samples, MI adversaries can maximize their attack success while minimizing the number of shadow models. First, our motivational experiments suggest memorization as the key property explaining disparate sample vulnerability to MIAs. We formalize this through a theoretical bound that connects MI advantage with memorization. Second, we show sample complexity bounds that connect the number of shadow models needed for MIAs with memorization. Lastly, we confirm our theoretical arguments with comprehensive experiments; by utilizing samples with high memorization scores, the adversary can (a) significantly improve its efficacy regardless of the MIA used, and (b) reduce the number of shadow models by nearly two orders of magnitude compared to state-of-the-art approaches.

Neal Mangaokar, Ashish Hooda, Jihye Choi et al.

Large language models (LLMs) are typically aligned to be harmless to humans. Unfortunately, recent work has shown that such models are susceptible to automated jailbreak attacks that induce them to generate harmful content. More recent LLMs often incorporate an additional layer of defense, a Guard Model, which is a second LLM that is designed to check and moderate the output response of the primary LLM. Our key contribution is to show a novel attack strategy, PRP, that is successful against several open-source (e.g., Llama 2) and closed-source (e.g., GPT 3.5) implementations of Guard Models. PRP leverages a two step prefix-based attack that operates by (a) constructing a universal adversarial prefix for the Guard Model, and (b) propagating this prefix to the response. We find that this procedure is effective across multiple threat models, including ones in which the adversary has no access to the Guard Model at all. Our work suggests that further advances are required on defenses and Guard Models before they can be considered effective.

Nils Palumbo, Ravi Mangal, Zifan Wang et al.

Mechanistic interpretability aims to reverse engineer the computation performed by a neural network in terms of its internal components. Although there is a growing body of research on mechanistic interpretation of neural networks, the notion of a mechanistic interpretation itself is often ad-hoc. Inspired by the notion of abstract interpretation from the program analysis literature that aims to develop approximate semantics for programs, we give a set of axioms that formally characterize a mechanistic interpretation as a description that approximately captures the semantics of the neural network under analysis in a compositional manner. We demonstrate the applicability of these axioms for validating mechanistic interpretations on an existing, well-known interpretability study as well as on a new case study involving a Transformer-based model trained to solve the well-known 2-SAT problem.

Irfan Ul Haq, Sergio Chica, Juan Caballero et al.

Malware lineage studies the evolutionary relationships among malware and has important applications for malware analysis. A persistent limitation of prior malware lineage approaches is to consider every input sample a separate malware version. This is problematic since a majority of malware are packed and the packing process produces many polymorphic variants (i.e., executables with different file hash) of the same malware version. Thus, many samples correspond to the same malware version and it is challenging to identify distinct malware versions from polymorphic variants. This problem does not manifest in prior malware lineage approaches because they work on synthetic malware, malware that are not packed, or packed malware for which unpackers are available. In this work, we propose a novel malware lineage approach that works on malware samples collected in the wild. Given a set of malware executables from the same family, for which no source code is available and which may be packed, our approach produces a lineage graph where nodes are versions of the family and edges describe the relationships between versions. To enable our malware lineage approach, we propose the first technique to identify the versions of a malware family and a scalable code indexing technique for determining shared functions between any pair of input samples. We have evaluated the accuracy of our approach on 13 open-source programs and have applied it to produce lineage graphs for 10 popular malware families. Our malware lineage graphs achieve on average a 26 times reduction from number of input samples to number of versions.

Xuandong Zhao, Sam Gunn, Miranda Christ et al. · berkeley, eth-zurich

As the outputs of generative AI (GenAI) techniques improve in quality, it becomes increasingly challenging to distinguish them from human-created content. Watermarking schemes are a promising approach to address the problem of distinguishing between AI and human-generated content. These schemes embed hidden signals within AI-generated content to enable reliable detection. While watermarking is not a silver bullet for addressing all risks associated with GenAI, it can play a crucial role in enhancing AI safety and trustworthiness by combating misinformation and deception. This paper presents a comprehensive overview of watermarking techniques for GenAI, beginning with the need for watermarking from historical and regulatory perspectives. We formalize the definitions and desired properties of watermarking schemes and examine the key objectives and threat models for existing approaches. Practical evaluation strategies are also explored, providing insights into the development of robust watermarking techniques capable of resisting various attacks. Additionally, we review recent representative works, highlight open challenges, and discuss potential directions for this emerging field. By offering a thorough understanding of watermarking in GenAI, this work aims to guide researchers in advancing watermarking methods and applications, and support policymakers in addressing the broader implications of GenAI.

Ashish Hooda, Mihai Christodorescu, Miltiadis Allamanis et al.

Large Language Models' success on text generation has also made them better at code generation and coding tasks. While a lot of work has demonstrated their remarkable performance on tasks such as code completion and editing, it is still unclear as to why. We help bridge this gap by exploring to what degree auto-regressive models understand the logical constructs of the underlying programs. We propose Counterfactual Analysis for Programming Concept Predicates (CACP) as a counterfactual testing framework to evaluate whether Large Code Models understand programming concepts. With only black-box access to the model, we use CACP to evaluate ten popular Large Code Models for four different programming concepts. Our findings suggest that current models lack understanding of concepts such as data flow and control flow.

Jihye Choi, Jayaram Raghuram, Yixuan Li et al.

Advancements in foundation models (FMs) have led to a paradigm shift in machine learning. The rich, expressive feature representations from these pre-trained, large-scale FMs are leveraged for multiple downstream tasks, usually via lightweight fine-tuning of a shallow fully-connected network following the representation. However, the non-interpretable, black-box nature of this prediction pipeline can be a challenge, especially in critical domains such as healthcare, finance, and security. In this paper, we explore the potential of Concept Bottleneck Models (CBMs) for transforming complex, non-interpretable foundation models into interpretable decision-making pipelines using high-level concept vectors. Specifically, we focus on the test-time deployment of such an interpretable CBM pipeline "in the wild", where the input distribution often shifts from the original training distribution. We first identify the potential failure modes of such a pipeline under different types of distribution shifts. Then we propose an adaptive concept bottleneck framework to address these failure modes, that dynamically adapts the concept-vector bank and the prediction layer based solely on unlabeled data from the target domain, without access to the source (training) dataset. Empirical evaluations with various real-world distribution shifts show that our adaptation method produces concept-based interpretations better aligned with the test data and boosts post-deployment accuracy by up to 28%, aligning the CBM performance with that of non-interpretable classification.

Jaiden Fairoze, Guillermo Ortiz-Jimenez, Mel Vecerik et al.

This work investigates the theoretical boundaries of creating publicly-detectable schemes to enable the provenance of watermarked imagery. Metadata-based approaches like C2PA provide unforgeability and public-detectability. ML techniques offer robust retrieval and watermarking. However, no existing scheme combines robustness, unforgeability, and public-detectability. In this work, we formally define such a scheme and establish its existence. Although theoretically possible, we find that at present, it is intractable to build certain components of our scheme without a leap in deep learning capabilities. We analyze these limitations and propose research directions that need to be addressed before we can practically realize robust and publicly-verifiable provenance.

Jihye Choi, Jinsung Yoon, Jiefeng Chen et al.

While Large Language Models (LLMs) have shown remarkable advancements in reasoning and tool use, they often fail to generate optimal, grounded solutions under complex constraints. Real-world travel planning exemplifies these challenges, evaluating agents' abilities to handle constraints that are explicit, implicit, and even evolving based on interactions with dynamic environments and user needs. In this paper, we present ATLAS, a general multi-agent framework designed to effectively handle such complex nature of constraints awareness in real-world travel planning tasks. ATLAS introduces a principled approach to address the fundamental challenges of constraint-aware planning through dedicated mechanisms for dynamic constraint management, iterative plan critique, and adaptive interleaved search. ATLAS demonstrates state-of-the-art performance on the TravelPlanner benchmark, improving the final pass rate from 23.3% to 44.4% over its best alternative. More importantly, our work is the first to demonstrate quantitative effectiveness on real-world travel planning tasks with live information search and multi-turn feedback. In this realistic setting, ATLAS showcases its superior overall planning performance, achieving an 84% final pass rate which significantly outperforms baselines including ReAct (59%) and a monolithic agent (27%).

Sarthak Choudhary, Divyam Anshumaan, Nils Palumbo et al.

LLM-integrated applications and agents are vulnerable to prompt injection attacks, in which adversaries embed malicious instructions within seemingly benign user inputs to manipulate the LLM's intended behavior. Recent defenses based on $\textit{known-answer detection}$ (KAD) have achieved near-perfect performance by using an LLM to classify inputs as clean or contaminated. In this work, we formally characterize the KAD framework and uncover a structural vulnerability in its design that invalidates its core security premise. We design a methodical adaptive attack, $\textit{DataFlip}$, to exploit this fundamental weakness. It consistently evades KAD defenses with detection rates as low as $1.5\%$ while reliably inducing malicious behavior with success rates of up to $88\%$, without needing white-box access to the LLM or any optimization procedures.

Mian Wei, Somesh Jha, David Page

The discovery of causal relationships is a foundational problem in artificial intelligence, statistics, epidemiology, economics, and beyond. While elegant theories exist for accurate causal discovery given infinite data, real-world applications are inherently resource-constrained. Effective methods for inferring causal relationships from observational data must perform well under finite data and time constraints, where "performing well" implies achieving high, though not perfect accuracy. In his seminal paper A Theory of the Learnable, Valiant highlighted the importance of resource constraints in supervised machine learning, introducing the concept of Probably Approximately Correct (PAC) learning as an alternative to exact learning. Inspired by Valiant's work, we propose the Probably Approximately Correct Causal (PACC) Discovery framework, which extends PAC learning principles to the causal field. This framework emphasizes both computational and sample efficiency for established causal methods such as propensity score techniques and instrumental variable approaches. Furthermore, we show that it can also provide theoretical guarantees for other widely used methods, such as the Self-Controlled Case Series (SCCS) method, which had previously lacked such guarantees.

Sarthak Choudhary, Nils Palumbo, Ashish Hooda et al.

Retrieval-augmented generation (RAG) systems are vulnerable to attacks that inject poisoned passages into the retrieved set, even at low corruption rates. We show that existing attacks are not designed to be stealthy, allowing reliable detection and mitigation. We formalize stealth using a distinguishability-based security game. If a few poisoned passages are designed to control the response, they must differentiate themselves from benign ones, inherently compromising stealth. This motivates the need for attackers to rigorously analyze intermediate signals involved in generation$\unicode{x2014}$such as attention patterns or next-token probability distributions$\unicode{x2014}$to avoid easily detectable traces of manipulation. Leveraging attention patterns, we propose a passage-level score$\unicode{x2014}$the Normalized Passage Attention Score$\unicode{x2014}$used by our Attention-Variance Filter algorithm to identify and filter potentially poisoned passages. This method mitigates existing attacks, improving accuracy by up to $\sim 20 \%$ over baseline defenses. To probe the limits of attention-based defenses, we craft stealthier adaptive attacks that obscure such traces, achieving up to $35 \%$ attack success rate, and highlight the challenges in improving stealth.

Zi Wang, Shiwei Weng, Mohannad Alhanahnah et al.

Large Language Models (LLMs) have exhibited remarkable capabilities across diverse domains, prompting investigations into their potential as generic reasoning engines. While recent studies have explored inference-time computation to enhance model performance on complex problems, current research lacks a formal framework to characterize the complexity of reasoning tasks. This study introduces the Predicate-Enumeration-Aggregation (PEA) framework, a formal approach to describe and solve a class of important reasoning tasks termed computational reasoning problems. The PEA framework decomposes these problems into predicate and enumeration components, using LLMs to synthesize programs based on specified predicates, enumeration, and aggregation rules. These synthesized programs are then executed to obtain solutions to the computational tasks. We demonstrate the framework's efficacy on benchmark tasks including Boolean satisfiability problems, game of $24$, and planning problems. Empirical evaluation reveals that PEA substantially enhances the performance of underlying models on benchmark computational problems, yielding an average accuracy improvement of approximately $50\%$, coupled with increased efficiency.

Jihye Choi, Sai Rahul Rachuri, Ke Wang et al.

Federated Learning (FL) enables collaborative model training while keeping client data private. However, exposing individual client updates makes FL vulnerable to reconstruction attacks. Secure aggregation mitigates such privacy risks but prevents the server from verifying the validity of each client update, creating a privacy-robustness tradeoff. Recent efforts attempt to address this tradeoff by enforcing checks on client updates using zero-knowledge proofs, but they support limited predicates and often depend on public validation data. We propose SLVR, a general framework that securely leverages clients' private data through secure multi-party computation. By utilizing clients' data, SLVR not only eliminates the need for public validation data, but also enables a wider range of checks for robustness, including cross-client accuracy validation. It also adapts naturally to distribution shifts in client data as it can securely refresh its validation data up-to-date. Our empirical evaluations show that SLVR improves robustness against model poisoning attacks, particularly outperforming existing methods by up to 50% under adaptive attacks. Additionally, SLVR demonstrates effective adaptability and stable convergence under various distribution shift scenarios.

Nils Palumbo, Yang Guo, Xi Wu et al.

Both transduction and rejection have emerged as important techniques for defending against adversarial perturbations. A recent work by Goldwasser et al. showed that rejection combined with transduction can give provable guarantees (for certain problems) that cannot be achieved otherwise. Nevertheless, under recent strong adversarial attacks, their work was shown to have low performance in a practical deep-learning setting. In this paper, we take a step towards realizing the promise of transduction+rejection in more realistic scenarios. Our key observation is that a novel application of a reduction technique by Tramèr, which was until now only used to demonstrate the vulnerability of certain defenses, can be used to actually construct effective defenses. Theoretically, we show that a careful application of this technique in the transductive setting can give significantly improved sample-complexity for robust generalization. Our theory guides us to design a new transductive algorithm for learning a selective model; extensive experiments using state of the art attacks show that our approach provides significantly better robust accuracy (81.6% on CIFAR-10 and 57.9% on CIFAR-100 under $l_\infty$ with budget 8/255) than existing techniques.

Jiefeng Chen, Jayaram Raghuram, Jihye Choi et al.

Recently, there is an emerging interest in adversarially training a classifier with a rejection option (also known as a selective classifier) for boosting adversarial robustness. While rejection can incur a cost in many applications, existing studies typically associate zero cost with rejecting perturbed inputs, which can result in the rejection of numerous slightly-perturbed inputs that could be correctly classified. In this work, we study adversarially-robust classification with rejection in the stratified rejection setting, where the rejection cost is modeled by rejection loss functions monotonically non-increasing in the perturbation magnitude. We theoretically analyze the stratified rejection setting and propose a novel defense method -- Adversarial Training with Consistent Prediction-based Rejection (CPR) -- for building a robust selective classifier. Experiments on image datasets demonstrate that the proposed method significantly outperforms existing methods under strong adaptive attacks. For instance, on CIFAR-10, CPR reduces the total robust loss (for different rejection losses) by at least 7.3% under both seen and unseen attacks.

Aiping Xiong, Chuhao Wu, Tianhao Wang et al.

Proper communication is key to the adoption and implementation of differential privacy (DP). However, a prior study found that laypeople did not understand the data perturbation processes of DP and how DP noise protects their sensitive personal information. Consequently, they distrusted the techniques and chose to opt out of participating. In this project, we designed explanative illustrations of three DP models (Central DP, Local DP, Shuffler DP) to help laypeople conceptualize how random noise is added to protect individuals' privacy and preserve group utility. Following pilot surveys and interview studies, we conducted two online experiments (N = 595) examining participants' comprehension, privacy and utility perception, and data-sharing decisions across the three DP models. Besides the comparisons across the three models, we varied the noise levels of each model. We found that the illustrations can be effective in communicating DP to the participants. Given an adequate comprehension of DP, participants preferred strong privacy protection for a certain type of data usage scenarios (i.e., commercial interests) at both the model level and the noise level. We also obtained empirical evidence showing participants' acceptance of the Shuffler DP model for data privacy protection. Our findings have implications for multiple stakeholders for user-centered deployments of differential privacy, including app developers, DP model developers, data curators, and online users.

Ashish Hooda, Neal Mangaokar, Ryan Feng et al.

Detecting diffusion-generated deepfake images remains an open problem. Current detection methods fail against an adversary who adds imperceptible adversarial perturbations to the deepfake to evade detection. In this work, we propose Disjoint Diffusion Deepfake Detection (D4), a deepfake detector designed to improve black-box adversarial robustness beyond de facto solutions such as adversarial training. D4 uses an ensemble of models over disjoint subsets of the frequency spectrum to significantly improve adversarial robustness. Our key insight is to leverage a redundancy in the frequency domain and apply a saliency partitioning technique to disjointly distribute frequency components across multiple models. We formally prove that these disjoint ensembles lead to a reduction in the dimensionality of the input subspace where adversarial deepfakes lie, thereby making adversarial deepfakes harder to find for black-box attacks. We then empirically validate the D4 method against several black-box attacks and find that D4 significantly outperforms existing state-of-the-art defenses applied to diffusion-generated deepfake detection. We also demonstrate that D4 provides robustness against adversarial deepfakes from unseen data distributions as well as unseen generative techniques.

Harrison Rosenberg, Robi Bhattacharjee, Kassem Fawaz et al.

Recent works have investigated the sample complexity necessary for fair machine learning. The most advanced of such sample complexity bounds are developed by analyzing multicalibration uniform convergence for a given predictor class. We present a framework which yields multicalibration error uniform convergence bounds by reparametrizing sample complexities for Empirical Risk Minimization (ERM) learning. From this framework, we demonstrate that multicalibration error exhibits dependence on the classifier architecture as well as the underlying data distribution. We perform an experimental evaluation to investigate the behavior of multicalibration error for different families of classifiers. We compare the results of this evaluation to multicalibration error concentration bounds. Our investigation provides additional perspective on both algorithmic fairness and multicalibration error convergence bounds. Given the prevalence of ERM sample complexity bounds, our proposed framework enables machine learning practitioners to easily understand the convergence behavior of multicalibration error for a myriad of classifier architectures.

Amrita Roy Chowdhury, Chuan Guo, Somesh Jha et al.

Federated learning (FL) enables clients to collaborate with a server to train a machine learning model. To ensure privacy, the server performs secure aggregation of updates from the clients. Unfortunately, this prevents verification of the well-formedness (integrity) of the updates as the updates are masked. Consequently, malformed updates designed to poison the model can be injected without detection. In this paper, we formalize the problem of ensuring \textit{both} update privacy and integrity in FL and present a new system, \textsf{EIFFeL}, that enables secure aggregation of \textit{verified} updates. \textsf{EIFFeL} is a general framework that can enforce \textit{arbitrary} integrity checks and remove malformed updates from the aggregate, without violating privacy. Our empirical evaluation demonstrates the practicality of \textsf{EIFFeL}. For instance, with $100$ clients and $10\%$ poisoning, \textsf{EIFFeL} can train an MNIST classification model to the same accuracy as that of a non-poisoned federated learner in just $2.4s$ per iteration.

Jiefeng Chen, Xi Wu, Yang Guo et al.

There has been emerging interest in using transductive learning for adversarial robustness (Goldwasser et al., NeurIPS 2020; Wu et al., ICML 2020; Wang et al., ArXiv 2021). Compared to traditional defenses, these defense mechanisms "dynamically learn" the model based on test-time input; and theoretically, attacking these defenses reduces to solving a bilevel optimization problem, which poses difficulty in crafting adaptive attacks. In this paper, we examine these defense mechanisms from a principled threat analysis perspective. We formulate and analyze threat models for transductive-learning based defenses, and point out important subtleties. We propose the principle of attacking model space for solving bilevel attack objectives, and present Greedy Model Space Attack (GMSA), an attack framework that can serve as a new baseline for evaluating transductive-learning based defenses. Through systematic evaluation, we show that GMSA, even with weak instantiations, can break previous transductive-learning based defenses, which were resilient to previous attacks, such as AutoAttack. On the positive side, we report a somewhat surprising empirical result of "transductive adversarial training": Adversarially retraining the model using fresh randomness at the test time gives a significant increase in robustness against attacks we consider.

Mohannad Alhanahnah, Rithik Jain, Vaibhav Rastogi et al.

Program debloating aims to enhance the performance and reduce the attack surface of bloated applications. Several techniques have been recently proposed to specialize programs. These approaches are either based on unsound strategies or demanding techniques, leading to unsafe results or a high overhead debloating process. In this paper, we address these limitations by applying partial-evaluation principles to generate specialized applications. Our approach relies on a simple observation that an application typically consists of configuration logic, followed by the main logic of the program. The configuration logic specifies what functionality in the main logic should be executed. LMCAS performs partial interpretation to capture a precise program state of the configuration logic based on the supplied inputs. LMCAS then applies partial-evaluation optimizations to generate a specialized program by propagating the constants in the captured partial state, eliminating unwanted code, and preserving the desired functionalities. Our evaluation of LMCAS on commonly used benchmarks and real-world applications shows that it successfully removes unwanted features while preserving the functionality and robustness of the deblated programs, runs faster than prior tools, and reduces the attack surface of specialized programs. LMCAS runs 1500x, 4.6x, and 1.2x faster than the state-of-the-art debloating tools CHISEL, RAZOR, and OCCAM, respectively; achieves 25% reduction in the binary size; reduces the attack surface of code-reuse attacks by removing 51.7% of the total gadgets and eliminating 83% of known CVE vulnerabilities

Nicholas Carlini, Sanjam Garg, Somesh Jha et al.

NeuraCrypt (Yara et al. arXiv 2021) is an algorithm that converts a sensitive dataset to an encoded dataset so that (1) it is still possible to train machine learning models on the encoded data, but (2) an adversary who has access only to the encoded dataset can not learn much about the original sensitive dataset. We break NeuraCrypt privacy claims, by perfectly solving the authors' public challenge, and by showing that NeuraCrypt does not satisfy the formal privacy definitions posed in the original paper. Our attack consists of a series of boosting steps that, coupled with various design flaws, turns a 1% attack advantage into a 100% complete break of the scheme.

Harrison Rosenberg, Brian Tang, Kassem Fawaz et al.

The proliferation of automated face recognition in the commercial and government sectors has caused significant privacy concerns for individuals. One approach to address these privacy concerns is to employ evasion attacks against the metric embedding networks powering face recognition systems: Face obfuscation systems generate imperceptibly perturbed images that cause face recognition systems to misidentify the user. Perturbed faces are generated on metric embedding networks, which are known to be unfair in the context of face recognition. A question of demographic fairness naturally follows: are there demographic disparities in face obfuscation system performance? We answer this question with an analytical and empirical exploration of recent face obfuscation systems. Metric embedding networks are found to be demographically aware: face embeddings are clustered by demographic. We show how this clustering behavior leads to reduced face obfuscation utility for faces in minority groups. An intuitive analytical model yields insight into these phenomena.

Jayaram Raghuram, Yijing Zeng, Dolores García Martí et al.

The problem of end-to-end learning of a communication system using an autoencoder -- consisting of an encoder, channel, and decoder modeled using neural networks -- has recently been shown to be a promising approach. A challenge faced in the practical adoption of this learning approach is that under changing channel conditions (e.g. a wireless link), it requires frequent retraining of the autoencoder in order to maintain a low decoding error rate. Since retraining is both time consuming and requires a large number of samples, it becomes impractical when the channel distribution is changing quickly. We propose to address this problem using a fast and sample-efficient (few-shot) domain adaptation method that does not change the encoder and decoder networks. Different from conventional training-time unsupervised or semi-supervised domain adaptation, here we have a trained autoencoder from a source distribution, that we want to adapt (at test time) to a target distribution using only a small labeled dataset and no unlabeled data. Our method focuses on a Gaussian mixture density network based channel model, and formulates its adaptation based on class and component-conditional affine transformations. The learned affine transformations are used to design an optimal input transformation at the decoder to compensate for the distribution shift, and effectively present to the decoder inputs close to the source distribution. Experiments on a real mmWave FPGA setup as well as a number of simulated distribution changes common to the wireless setting demonstrate the effectiveness of our method at adaptation using very small number of target domain samples.

Jiefeng Chen, Frederick Liu, Besim Avci et al.

When a deep learning model is deployed in the wild, it can encounter test data drawn from distributions different from the training data distribution and suffer drop in performance. For safe deployment, it is essential to estimate the accuracy of the pre-trained model on the test data. However, the labels for the test inputs are usually not immediately available in practice, and obtaining them can be expensive. This observation leads to two challenging tasks: (1) unsupervised accuracy estimation, which aims to estimate the accuracy of a pre-trained classifier on a set of unlabeled test inputs; (2) error detection, which aims to identify mis-classified test inputs. In this paper, we propose a principled and practically effective framework that simultaneously addresses the two tasks. The proposed framework iteratively learns an ensemble of models to identify mis-classified data points and performs self-training to improve the ensemble with the identified points. Theoretical analysis demonstrates that our framework enjoys provable guarantees for both accuracy estimation and error detection under mild conditions readily satisfied by practical deep learning models. Along with the framework, we proposed and experimented with two instantiations and achieved state-of-the-art results on 59 tasks. For example, on iWildCam, one instantiation reduces the estimation error for unsupervised accuracy estimation by at least 70% and improves the F1 score for error detection by at least 4.7% compared to existing methods.

Jiefeng Chen, Yang Guo, Xi Wu et al.

There has been emerging interest to use transductive learning for adversarial robustness (Goldwasser et al., NeurIPS 2020; Wu et al., ICML 2020). Compared to traditional "test-time" defenses, these defense mechanisms "dynamically retrain" the model based on test time input via transductive learning; and theoretically, attacking these defenses boils down to bilevel optimization, which seems to raise the difficulty for adaptive attacks. In this paper, we first formalize and analyze modeling aspects of transductive robustness. Then, we propose the principle of attacking model space for solving bilevel attack objectives, and present an instantiation of the principle which breaks previous transductive defenses. These attacks thus point to significant difficulties in the use of transductive learning to improve adversarial robustness. To this end, we present new theoretical and empirical evidence in support of the utility of transductive learning.

Casey Meehan, Amrita Roy Chowdhury, Kamalika Chaudhuri et al.

ldp deployments are vulnerable to inference attacks as an adversary can link the noisy responses to their identity and subsequently, auxiliary information using the order of the data. An alternative model, shuffle DP, prevents this by shuffling the noisy responses uniformly at random. However, this limits the data learnability -- only symmetric functions (input order agnostic) can be learned. In this paper, we strike a balance and show that systematic shuffling of the noisy responses can thwart specific inference attacks while retaining some meaningful data learnability. To this end, we propose a novel privacy guarantee, d-sigma-privacy, that captures the privacy of the order of a data sequence. d-sigma-privacy allows tuning the granularity at which the ordinal information is maintained, which formalizes the degree the resistance to inference attacks trading it off with data learnability. Additionally, we propose a novel shuffling mechanism that can achieve \name-privacy and demonstrate the practicality of our mechanism via evaluation on real-world datasets.

Varun Chandrasekaran, Darren Edge, Somesh Jha et al.

Making evidence based decisions requires data. However for real-world applications, the privacy of data is critical. Using synthetic data which reflects certain statistical properties of the original data preserves the privacy of the original data. To this end, prior works utilize differentially private data release mechanisms to provide formal privacy guarantees. However, such mechanisms have unacceptable privacy vs. utility trade-offs. We propose incorporating causal information into the training process to favorably modify the aforementioned trade-off. We theoretically prove that generative models trained with additional causal knowledge provide stronger differential privacy guarantees. Empirically, we evaluate our solution comparing different models based on variational auto-encoders (VAEs), and show that causal information improves resilience to membership inference, with improvements in downstream utility.

Washington Garcia, Pin-Yu Chen, Somesh Jha et al.

Designing deep networks robust to adversarial examples remains an open problem. Likewise, recent zeroth order hard-label attacks on image classification models have shown comparable performance to their first-order, gradient-level alternatives. It was recently shown in the gradient-level setting that regular adversarial examples leave the data manifold, while their on-manifold counterparts are in fact generalization errors. In this paper, we argue that query efficiency in the zeroth-order setting is connected to an adversary's traversal through the data manifold. To explain this behavior, we propose an information-theoretic argument based on a noisy manifold distance oracle, which leaks manifold information through the adversary's gradient estimate. Through numerical experiments of manifold-gradient mutual information, we show this behavior acts as a function of the effective problem dimensionality and number of training points. On real-world datasets and multiple zeroth-order attacks using dimension-reduction, we observe the same universal behavior to produce samples closer to the data manifold. This results in up to two-fold decrease in the manifold distance measure, regardless of the model robustness. Our results suggest that taking the manifold-gradient mutual information into account can thus inform better robust model design in the future, and avoid leakage of the sensitive data manifold.

Thomas Kobber Panum, Zi Wang, Pengyu Kan et al.

Deep Metric Learning (DML), a widely-used technique, involves learning a distance metric between pairs of samples. DML uses deep neural architectures to learn semantic embeddings of the input, where the distance between similar examples is small while dissimilar ones are far apart. Although the underlying neural networks produce good accuracy on naturally occurring samples, they are vulnerable to adversarially-perturbed samples that reduce performance. We take a first step towards training robust DML models and tackle the primary challenge of the metric losses being dependent on the samples in a mini-batch, unlike standard losses that only depend on the specific input-output pair. We analyze this dependence effect and contribute a robust optimization formulation. Using experiments on three commonly-used DML datasets, we demonstrate 5-76 fold increases in adversarial accuracy, and outperform an existing DML model that sought out to be robust.