Machine learning (ML) models are susceptible to various security, privacy, and fairness risks. Adversaries with different characteristics (i.e., objectives, knowledge, and capabilities) can collude by executing one attack to amplify others. Existing work lacks a systematic framework to explore collusion among adversaries, and to study the implications of the adversaries' characteristics. We present a framework covering collusion (a) between train- and inference-time adversaries, and (b) among inference-time adversaries. Our framework accounts for factors enabling collusion between adversaries. We propose a guideline to conjecture about the potential for collusion using enabling factors. We use it to explain prior work, conjecture about unexplored collusions, and empirically validate five such cases. Finally, we discuss how adversaries' characteristics influence the potential for collusion.

Chatbot service providers (e.g., OpenAI) rely on tiered subscription plans to generate revenue, offering black-box access to basic models for free users and advanced models to paying subscribers. However, this approach is unprofitable and inflexible. A pay-to-unlock scheme for premium features (e.g., math, coding) offers a more sustainable alternative. Enabling such a scheme requires a feature-locking technique (FLoTE) that is (i) *effective* in refusing locked features, (ii) *utility-preserving* for unlocked features, (iii) *robust* against evasion or unauthorized credential sharing, and (iv) *scalable* to multiple features and clients. Existing FLoTEs (e.g., password-locked models) fail to meet these criteria. To fill this gap, we present Locket, a more *robust and scalable* FLoTE to enable pay-to-unlock schemes. We develop a framework for adversarial training and merging of feature-locking *adapters*, which enables Locket to selectively disable specific features of a model. Evaluation shows that Locket is effective (100% refusal rate), utility-preserving ( ≤ 7% utility degradation), robust ( ≤ 5% attack success rate), and scalable to multiple features and clients.

Machine learning property attestations allow provers (e.g., model providers or owners) to attest properties of their models/datasets to verifiers (e.g., regulators, customers), enabling accountability towards regulations and policies. But, current approaches do not support generative models or large datasets. We present PAL*M, a property attestation framework for large generative models, illustrated using large language models. PAL*M defines properties across training and inference, leverages confidential virtual machines with security-aware GPUs for coverage of CPU-GPU operations, and proposes using incremental multiset hashing over memory-mapped datasets to efficiently track their integrity. We implement PAL*M on Intel TDX and NVIDIA H100, showing it is efficient, scalable, versatile, and secure.

Language models (LMs) may memorize personally identifiable information (PII) from training data, enabling adversaries to extract it during inference. Existing defense mechanisms such as differential privacy (DP) reduce this leakage, but incur large drops in utility. Based on a comprehensive study using circuit discovery to identify the computational circuits responsible PII leakage in LMs, we hypothesize that specific PII leakage circuits in LMs should be responsible for this behavior. Therefore, we propose PATCH (Privacy-Aware Targeted Circuit PatcHing), a novel approach that first identifies and subsequently directly edits PII circuits to reduce leakage. PATCH achieves better privacy-utility trade-off than existing defenses, e.g., reducing recall of PII leakage from LMs by up to 65

Diffusion based text-to-image models are trained on large datasets scraped from the Internet, potentially containing unacceptable concepts (e.g., copyright-infringing or unsafe). We need concept removal techniques (CRTs) which are i) effective in preventing the generation of images with unacceptable concepts, ii) utility-preserving on acceptable concepts, and, iii) robust against evasion with adversarial prompts. No prior CRT satisfies all these requirements simultaneously. We introduce Espresso, the first robust concept filter based on Contrastive Language-Image Pre-Training (CLIP). We identify unacceptable concepts by using the distance between the embedding of a generated image to the text embeddings of both unacceptable and acceptable concepts. This lets us fine-tune for robustness by separating the text embeddings of unacceptable and acceptable concepts while preserving utility. We present a pipeline to evaluate various CRTs to show that Espresso is more effective and robust than prior CRTs, while retaining utility.

Machine learning community is discovering Contextual Integrity (CI) as a useful framework to assess the privacy implications of large language models (LLMs). This is an encouraging development. The CI theory emphasizes sharing information in accordance with privacy norms and can bridge the social, legal, political, and technical aspects essential for evaluating privacy in LLMs. However, this is also a good point to reflect on use of CI for LLMs. This position paper argues that existing literature inadequately applies CI for LLMs without embracing the theory's fundamental tenets. Inadequate applications of CI could lead to incorrect conclusions and flawed privacy-preserving designs. We clarify the four fundamental tenets of CI theory, systematize prior work on whether they deviate from these tenets, and highlight overlooked issues in experimental hygiene for LLMs (e.g., prompt sensitivity, positional bias).

Machine learning (ML) defenses protect against various risks to security, privacy, and fairness. Real-life models need simultaneous protection against multiple different risks which necessitates combining multiple defenses. But combining defenses with conflicting interactions in an ML model can be ineffective, incurring a significant drop in the effectiveness of one or more defenses being combined. Practitioners need a way to determine if a given combination can be effective. Experimentally identifying effective combinations can be time-consuming and expensive, particularly when multiple defenses need to be combined. We need an inexpensive, easy-to-use combination technique to identify effective combinations. Ideally, a combination technique should be (a) accurate (correctly identifies whether a combination is effective or not), (b) scalable (allows combining multiple defenses), (c) non-invasive (requires no change to the defenses being combined), and (d) general (is applicable to different types of defenses). Prior works have identified several ad-hoc techniques but none satisfy all the requirements above. We propose a principled combination technique, Def, to identify effective defense combinations. Defmeets all requirements, achieving 90

Regulations increasingly call for various assurances from machine learning (ML) model providers about their training data, training process, and model behavior. For better transparency, industry (e.g., Huggingface and Google) has adopted model cards and datasheets to describe various properties of training datasets and models. In the same vein, we introduce the notion of inference cards to describe the properties of a given inference (e.g., binding of the output to the model and its corresponding input). We coin the term ML property cards to collectively refer to these various types of cards. To prevent a malicious model provider from including false information in ML property cards, they need to be verifiable. We show how to construct verifiable ML property cards using property attestation, technical mechanisms by which a prover (e.g., a model provider) can attest to various ML properties to a verifier (e.g., an auditor). Since prior attestation mechanisms based purely on cryptography are often narrowly focused (lacking versatility) and inefficient, we need an efficient mechanism to attest different types of properties across the entire ML model pipeline. Emerging widespread support for confidential computing has made it possible to run and even train models inside hardware-assisted trusted execution environments (TEEs), which provide highly efficient attestation mechanisms. We propose Laminator, which uses TEEs to provide the first framework for verifiable ML property cards via hardware-assisted ML property attestations. Laminator is efficient in terms of overhead, scalable to large numbers of verifiers, and versatile with respect to the properties it can prove during training or inference.

Machine learning (ML) models are susceptible to various risks to security, privacy, and fairness. Most defenses are designed to protect against each risk individually (intended interactions) but can inadvertently affect susceptibility to other unrelated risks (unintended interactions). We introduce Amulet, the first Python library for evaluating both intended and unintended interactions among ML defenses and risks. Amulet is comprehensive by including representative attacks, defenses, and metrics; extensible to new modules due to its modular design; consistent with a user-friendly API template for inputs and outputs; and applicable for evaluating novel interactions. By satisfying all four properties, Amulet offers a unified foundation for studying how defenses interact, enabling the first systematic evaluation of unintended interactions across multiple risks.

The success of machine learning (ML) has been accompanied by increased concerns about its trustworthiness. Several jurisdictions are preparing ML regulatory frameworks. One such concern is ensuring that model training data has desirable distributional properties for certain sensitive attributes. For example, draft regulations indicate that model trainers are required to show that training datasets have specific distributional properties, such as reflecting diversity of the population. We propose the notion of property attestation allowing a prover (e.g., model trainer) to demonstrate relevant distributional properties of training data to a verifier (e.g., a customer) without revealing the data. We present an effective hybrid property attestation combining property inference with cryptographic mechanisms.

As large language models (LLMs) are integrated into sociotechnical systems, it is crucial to examine the privacy biases they exhibit. We define privacy bias as the appropriateness value of information flows in responses from LLMs. A deviation between privacy biases and expected values, referred to as privacy bias delta, may indicate privacy violations. As an auditing metric, privacy bias can help (a) model trainers evaluate the ethical and societal impact of LLMs, (b) service providers select context-appropriate LLMs, and (c) policymakers assess the appropriateness of privacy biases in deployed LLMs. We formulate and answer a novel research question: how can we reliably examine privacy biases in LLMs and the factors that influence them? We present a novel approach for assessing privacy biases using a contextual integrity-based methodology to evaluate the responses from various LLMs. Our approach accounts for the sensitivity of responses across prompt variations, which hinders the evaluation of privacy biases. Finally, we investigate how privacy biases are affected by model capacities and optimizations.

Machine learning (ML) models have been adopted for applications with high-stakes decision-making like healthcare and criminal justice. To ensure trustworthy ML models, the new AI regulations (e.g., AI Act) have established several pillars such as privacy, safety and fairness that model design must take into account. Designing such models requires an understanding of the interactions between fairness definitions with different notions of privacy. Specifically, the interaction of group fairness (i.e., protection against discriminatory behaviour across demographic subgroups) with attribute privacy (i.e., resistance to attribute inference attacks—AIA s), has not been comprehensively studied. In this paper, we study in depth, both theoretically and empirically, the alignment of group fairness with attribute privacy in a blackbox setting. We first propose AdaptAIA, which outperforms existing AIA s on real-world datasets with class imbalances in sensitive attributes. We then show that group fairness theoretically bounds the success of AdaptAIA, which depends on the choice of fairness metrics (e.g., demographic parity or equalized odds). Through our empirical study, we show that attribute privacy can be achieved from group fairness at no additional cost other than the already existing trade-off with utility. Our work has several implications: i) group fairness acts as a defense against AIA s, which is currently lacking ii) practitioners do not need to explicitly train models for both fairness and privacy to meet regulatory requirements iii) AdaptAIA can be used for blackbox auditing of group fairness.

Machine learning (ML) models cannot neglect risks to security, privacy, and fairness. Several defenses have been proposed to mitigate such risks. When a defense is effective in mitigating one risk, it may correspond to increased or decreased susceptibility to other risks. Existing research lacks an effective framework to recognize and explain these unintended interactions. We present such a framework, based on the conjecture that overfitting and memorization underlie unintended interactions. We survey existing literature on unintended interactions, accommodating them within our framework. We use our framework to conjecture on two previously unexplored interactions, and empirically validate our conjectures.

Graph neural networks (GNNs) have emerged as a state-of-the-art approach to model and draw inferences from large scale graph-structured data in various application settings such as social networking. The primary goal of a GNN is to learn an embedding for each graph node in a dataset that encodes both the node features and the local graph structure around the node. Embeddings generated by a GNN for a graph node are unique to that GNN. Prior work has shown that GNNs are prone to model extraction attacks. Model extraction attacks and defenses have been explored extensively in other non-graph settings. While detecting or preventing model extraction appears to be difficult, deterring them via effective ownership verification techniques offer a potential defense. In non-graph settings, fingerprinting models, or the data used to build them, have shown to be a promising approach toward ownership verification. We present GrOVe, a state-of-the-art GNN model fingerprinting scheme that, given a target model and a suspect model, can reliably determine if the suspect model was trained independently of the target model or if it is a surrogate of the target model obtained via model extraction. We show that GrOVe can distinguish between surrogate and independent models even when the independent model uses the same training dataset and architecture as the original target model. Using six benchmark datasets and three model architectures, we show that consistently achieves low false-positive and false-negative rates. We demonstrate that is robust against known fingerprint evasion techniques while remaining computationally efficient.

Private computation, which includes techniques like multi-party computation and private query execution, holds great promise for enabling organizations to analyze data they and their partners hold while maintaining data subjects' privacy. Despite recent interest in communicating about differential privacy, end users' perspectives on private computation have not previously been studied. To fill this gap, we conducted 22 semi-structured interviews investigating users' understanding of, and expectations for, private computation over data about them. Interviews centered on four concrete data-analysis scenarios (e.g., ad conversion analysis), each with a variant that did not use private computation and another that did. While participants struggled with abstract definitions of private computation, they found the concrete scenarios enlightening and plausible even though we did not explain the complex cryptographic underpinnings. Private computation increased participants' acceptance of data sharing, but not unconditionally; the purpose of data sharing and analysis was the primary driver of their attitudes. Through collective activities, participants emphasized the importance of detailing the purpose of a computation and clarifying that inputs to private computation are not shared across organizations when describing private computation to end users.

Embedded systems demand on-device processing of data using Neural Networks (NNs) while conforming to the memory, power and computation constraints, leading to an efficiency and accuracy tradeoff. To bring NNs to edge devices, several optimizations such as model compression through pruning, quantization, and off-the-shelf architectures with efficient design have been extensively adopted. These algorithms when deployed to real world sensitive applications, requires to resist inference attacks to protect privacy of users training data. However, resistance against inference attacks is not accounted for designing NN models for IoT. In this work, we analyse the three-dimensional privacy-accuracy-efficiency tradeoff in NNs for IoT devices and propose Gecko training methodology where we explicitly add resistance to private inferences as a design objective. We optimize the inference-time memory, computation, and power constraints of embedded devices as a criterion for designing NN architecture while also preserving privacy. We choose quantization as design choice for highly efficient and private models. This choice is driven by the observation that compressed models leak more information compared to baseline models while off-the-shelf efficient architectures indicate poor efficiency and privacy tradeoff. We show that models trained using Gecko methodology are comparable to prior defences against black-box membership attacks in terms of accuracy and privacy while providing efficiency.

Machine learning (ML) models have been deployed for high-stakes applications. Due to class imbalance in the sensitive attribute observed in the datasets, ML models are unfair on minority subgroups identified by a sensitive attribute, such as race and sex. In-processing fairness algorithms ensure model predictions are independent of sensitive attribute. Furthermore, ML models are vulnerable to attribute inference attacks where an adversary can identify the values of sensitive attribute by exploiting their distinguishable model predictions. Despite privacy and fairness being important pillars of trustworthy ML, the privacy risk introduced by fairness algorithms with respect to attribute leakage has not been studied. We identify attribute inference attacks as an effective measure for auditing blackbox fairness algorithms to enable model builder to account for privacy and fairness in the model design. We proposed Dikaios, a privacy auditing tool for fairness algorithms for model builders which leveraged a new effective attribute inference attack that account for the class imbalance in sensitive attributes through an adaptive prediction threshold. We evaluated Dikaios to perform a privacy audit of two in-processing fairness algorithms over five datasets. We show that our attribute inference attacks with adaptive prediction threshold significantly outperform prior attacks. We highlighted the limitations of in-processing fairness algorithms to ensure indistinguishable predictions across different values of sensitive attributes. Indeed, the attribute privacy risk of these in-processing fairness schemes is highly variable according to the proportion of the sensitive attributes in the dataset. This unpredictable effect of fairness mechanisms on the attribute privacy risk is an important limitation on their utilization which has to be accounted by the model builder.

Model explanations provide transparency into a trained machine learning model's blackbox behavior to a model builder. They indicate the influence of different input attributes to its corresponding model prediction. The dependency of explanations on input raises privacy concerns for sensitive user data. However, current literature has limited discussion on privacy risks of model explanations. We focus on the specific privacy risk of attribute inference attack wherein an adversary infers sensitive attributes of an input (e.g., Race and Sex) given its model explanations. We design the first attribute inference attack against model explanations in two threat models where model builder either (a) includes the sensitive attributes in training data and input or (b) censors the sensitive attributes by not including them in the training data and input. We evaluate our proposed attack on four benchmark datasets and four state-of-the-art algorithms. We show that an adversary can successfully infer the value of sensitive attributes from explanations in both the threat models accurately. Moreover, the attack is successful even by exploiting only the explanations corresponding to sensitive attributes. These suggest that our attack is effective against explanations and poses a practical threat to data privacy. On combining the model predictions (an attack surface exploited by prior attacks) with explanations, we note that the attack success does not improve. Additionally, the attack success on exploiting model explanations is better compared to exploiting only model predictions. These suggest that model explanations are a strong attack surface to exploit for an adversary.

Machine learning models are typically made available to potential client users via inference APIs. Model extraction attacks occur when a malicious client uses information gleaned from queries to the inference API of a victim model $F_V$ to build a surrogate model $F_A$ that has comparable functionality. Recent research has shown successful model extraction attacks against image classification, and NLP models. In this paper, we show the first model extraction attack against real-world generative adversarial network (GAN) image translation models. We present a framework for conducting model extraction attacks against image translation models, and show that the adversary can successfully extract functional surrogate models. The adversary is not required to know $F_V$'s architecture or any other information about it beyond its intended image translation task, and queries $F_V$'s inference interface using data drawn from the same domain as the training data for $F_V$. We evaluate the effectiveness of our attacks using three different instances of two popular categories of image translation: (1) Selfie-to-Anime and (2) Monet-to-Photo (image style transfer), and (3) Super-Resolution (super resolution). Using standard performance metrics for GANs, we show that our attacks are effective in each of the three cases -- the differences between $F_V$ and $F_A$, compared to the target are in the following ranges: Selfie-to-Anime: FID $13.36-68.66$, Monet-to-Photo: FID $3.57-4.40$, and Super-Resolution: SSIM: $0.06-0.08$ and PSNR: $1.43-4.46$. Furthermore, we conducted a large scale (125 participants) user study on Selfie-to-Anime and Monet-to-Photo to show that human perception of the images produced by the victim and surrogate models can be considered equivalent, within an equivalence bound of Cohen's $d=0.3$.

Deep Learning Accelerators and Neuromorphic hardware, used in many real-time safety-critical applications, are prone to faults that manifest in the form of errors in Neural Networks. Fault Tolerance in Neural Networks is a critical attribute for applications that require reliable computation for long duration such as IoT and mobile devices. The inherent fault tolerance of Neural Networks can be improved with regularization, however, the current techniques exhibit a trade-off between generalization and classification accuracy. To this extent, in this work, a Neural Network is modelled as two distinct functional components: a Feature Extractor with an unsupervised learning objective and a Fully Connected Classifier with a supervised learning objective. Traditional approaches to train the entire network using a single supervised learning objective are insufficient to achieve the objectives of the individual functional goals optimally. In this work, a novel two phase framework with multi-criteria objective function combining unsupervised training of the Feature Extractor followed by supervised training of the Classifier Network is proposed. In the Phase I, the unsupervised training of the Feature Extractor is modeled using two games solved simultaneously in the presence of Neural Networks with conflicting objectives. The first game with a generative model, trains the Feature Extractor to generate robust features for the input image by minimizing a reconstruction loss between the input and reconstructed image. The second game with a binary classification network, updates the Feature Extractor to smoothen the feature space and match with a prior Gaussian distribution. In Phase II, the resultant Feature Extractor, which is strongly regularized, is combined with the Fully Connected Classifier for fine-tuning on the classification task. The proposed two phase training algorithm is evaluated on four architectures with varying model complexity on standard image classification datasets: FashionMNIST and CIFAR10. The proposed framework is scalable and independent of the network architecture that provides superior tolerance to stuck at "0" faults as compared to existing regularization functions without loss in classification accuracy.