We introduce Epsilon*, a new privacy metric for measuring the privacy risk of a single model instance prior to, during, or after deployment of privacy mitigation strategies. The metric requires only black-box access to model predictions, does not require training data re-sampling or model re-training, and can be used to measure the privacy risk of models not trained with differential privacy. Epsilon* is a function of true positive and false positive rates in a hypothesis test used by an adversary in a membership inference attack. We distinguish between quantifying the privacy loss of a trained model instance, which we refer to as empirical privacy, and quantifying the privacy loss of the training mechanism which produces this model instance. Existing approaches in the privacy auditing literature provide lower bounds for the latter, while our metric provides an empirical lower bound for the former by relying on an (${\epsilon}$, ${\delta}$)-type of quantification of the privacy of the trained model instance. We establish a relationship between these lower bounds and show how to implement Epsilon* to avoid numerical and noise amplification instability. We further show in experiments on benchmark public data sets that Epsilon* is sensitive to privacy risk mitigation by training with differential privacy (DP), where the value of Epsilon* is reduced by up to 800% compared to the Epsilon* values of non-DP trained baseline models. This metric allows privacy auditors to be independent of model owners, and enables visualizing the privacy-utility landscape to make informed decisions regarding the trade-offs between model privacy and utility.
Vertical federated learning (VFL) enables multiple parties with disjoint features of a common user set to train a machine learning model without sharing their private data. Tree-based models have become prevalent in VFL due to their interpretability and efficiency. However, the vulnerability of tree-based VFL has not been sufficiently investigated. In this study, we first introduce a novel label inference attack, ID2Graph, which utilizes the sets of record IDs assigned to each node (i.e., instance space)to deduce private training labels. ID2Graph attack generates a graph structure from training samples, extracts communities from the graph, and clusters the local dataset using community information. To counteract label leakage from the instance space, we propose two effective defense mechanisms, Grafting-LDP, which improves the utility of label differential privacy with post-processing, and andID-LMID, which focuses on mutual information regularization. Comprehensive experiments on various datasets reveal that ID2Graph presents significant risks to tree-based models such as RandomForest and XGBoost. Further evaluations of these benchmarks demonstrate that our defense methods effectively mitigate label leakage in such instances
Several membership inference (MI) attacks have been proposed to audit a target DNN. Given a set of subjects, MI attacks tell which subjects the target DNN has seen during training. This work focuses on the post-training MI attacks emphasizing high confidence membership detection -- True Positive Rates (TPR) at low False Positive Rates (FPR). Current works in this category -- likelihood ratio attack (LiRA) and enhanced MI attack (EMIA) -- only perform well on complex datasets (e.g., CIFAR-10 and Imagenet) where the target DNN overfits its train set, but perform poorly on simpler datasets (0% TPR by both attacks on Fashion-MNIST, 2% and 0% TPR respectively by LiRA and EMIA on MNIST at 1% FPR). To address this, firstly, we unify current MI attacks by presenting a framework divided into three stages -- preparation, indication and decision. Secondly, we utilize the framework to propose two novel attacks: (1) Adversarial Membership Inference Attack (AMIA) efficiently utilizes the membership and the non-membership information of the subjects while adversarially minimizing a novel loss function, achieving 6% TPR on both Fashion-MNIST and MNIST datasets; and (2) Enhanced AMIA (E-AMIA) combines EMIA and AMIA to achieve 8% and 4% TPRs on Fashion-MNIST and MNIST datasets respectively, at 1% FPR. Thirdly, we introduce two novel augmented indicators that positively leverage the loss information in the Gaussian neighborhood of a subject. This improves TPR of all four attacks on average by 2.5% and 0.25% respectively on Fashion-MNIST and MNIST datasets at 1% FPR. Finally, we propose simple, yet novel, evaluation metric, the running TPR average (RTA) at a given FPR, that better distinguishes different MI attacks in the low FPR region. We also show that AMIA and E-AMIA are more transferable to the unknown DNNs (other than the target DNN) and are more robust to DP-SGD training as compared to LiRA and EMIA.
Generative diffusion models, including Stable Diffusion and Midjourney, can generate visually appealing, diverse, and high-resolution images for various applications. These models are trained on billions of internet-sourced images, raising significant concerns about the potential unauthorized use of copyright-protected images. In this paper, we examine whether it is possible to determine if a specific image was used in the training set, a problem known in the cybersecurity community and referred to as a membership inference attack. Our focus is on Stable Diffusion, and we address the challenge of designing a fair evaluation framework to answer this membership question. We propose a methodology to establish a fair evaluation setup and apply it to Stable Diffusion, enabling potential extensions to other generative models. Utilizing this evaluation setup, we execute membership attacks (both known and newly introduced). Our research reveals that previously proposed evaluation setups do not provide a full understanding of the effectiveness of membership inference attacks. We conclude that the membership inference attack remains a significant challenge for large diffusion models (often deployed as black-box systems), indicating that related privacy and copyright issues will persist in the foreseeable future.
Collaborative-learning-based recommender systems, such as those employing Federated Learning (FL) and Gossip Learning (GL), allow users to train models while keeping their history of liked items on their devices. While these methods were seen as promising for enhancing privacy, recent research has shown that collaborative learning can be vulnerable to various privacy attacks. In this paper, we propose a novel attack called Community Inference Attack (CIA), which enables an adversary to identify community members based on a set of target items. What sets CIA apart is its efficiency: it operates at low computational cost by eliminating the need for training surrogate models. Instead, it uses a comparison-based approach, inferring sensitive information by comparing users' models rather than targeting any specific individual model. To evaluate the effectiveness of CIA, we conduct experiments on three real-world recommendation datasets using two recommendation models under both Federated and Gossip-like settings. The results demonstrate that CIA can be up to 10 times more accurate than random guessing. Additionally, we evaluate two mitigation strategies: Differentially Private Stochastic Gradient Descent (DP-SGD) and a Share less policy, which involves sharing fewer, less sensitive model parameters. Our findings suggest that the Share less strategy offers a better privacy-utility trade-off, especially in GL.
Membership inference attack is one of the most popular privacy attacks in machine learning, which aims to predict whether a given sample was contained in the target model's training set. Label-only membership inference attack is a variant that exploits sample robustness and attracts more attention since it assumes a practical scenario in which the adversary only has access to the predicted labels of the input samples. However, since the decision boundary distance, which measures robustness, is strongly affected by the random initial image, the adversary may get opposite results even for the same input samples. In this paper, we propose a new attack method, called muti-class adaptive membership inference attack in the label-only setting. All decision boundary distances for all target classes have been traversed in the early attack iterations, and the subsequent attack iterations continue with the shortest decision boundary distance to obtain a stable and optimal decision boundary distance. Instead of using a single boundary distance, the relative boundary distance between samples and neighboring points has also been employed as a new membership score to distinguish between member samples inside the training set and nonmember samples outside the training set. Experiments show that previous label-only membership inference attacks using the untargeted HopSkipJump algorithm fail to achieve optimal decision bounds in more than half of the samples, whereas our multi-targeted HopSkipJump algorithm succeeds in almost all samples. In addition, extensive experiments show that our multi-class adaptive MIA outperforms current label-only membership inference attacks in the CIFAR10, and CIFAR100 datasets, especially for the true positive rate at low false positive rates metric.
Graph neural networks (GNNs) have shown promising results on real-life datasets and applications, including healthcare, finance, and education. However, recent studies have shown that GNNs are highly vulnerable to attacks such as membership inference attack and link reconstruction attack. Surprisingly, attribute inference attacks has received little attention. In this paper, we initiate the first investigation into attribute inference attack where an attacker aims to infer the sensitive user attributes based on her public or non-sensitive attributes. We ask the question whether black-box attribute inference attack constitutes a significant privacy risk for graph-structured data and their corresponding GNN model. We take a systematic approach to launch the attacks by varying the adversarial knowledge and assumptions. Our findings reveal that when an attacker has black-box access to the target model, GNNs generally do not reveal significantly more information compared to missing value estimation techniques. Code is available.
Recently, diffusion models have achieved remarkable success in generating tasks, including image and audio generation. However, like other generative models, diffusion models are prone to privacy issues. In this paper, we propose an efficient query-based membership inference attack (MIA), namely Proximal Initialization Attack (PIA), which utilizes groundtruth trajectory obtained by $\epsilon$ initialized in $t=0$ and predicted point to infer memberships. Experimental results indicate that the proposed method can achieve competitive performance with only two queries on both discrete-time and continuous-time diffusion models. Moreover, previous works on the privacy of diffusion models have focused on vision tasks without considering audio tasks. Therefore, we also explore the robustness of diffusion models to MIA in the text-to-speech (TTS) task, which is an audio generation task. To the best of our knowledge, this work is the first to study the robustness of diffusion models to MIA in the TTS task. Experimental results indicate that models with mel-spectrogram (image-like) output are vulnerable to MIA, while models with audio output are relatively robust to MIA. {Code is available at \url{https://github.com/kong13661/PIA}}.
Large language models (LLMs) are excellent in-context learners. However, the sensitivity of data contained in prompts raises privacy concerns. Our work first shows that these concerns are valid: we instantiate a simple but highly effective membership inference attack against the data used to prompt LLMs. To address this vulnerability, one could forego prompting and resort to fine-tuning LLMs with known algorithms for private gradient descent. However, this comes at the expense of the practicality and efficiency offered by prompting. Therefore, we propose to privately learn to prompt. We first show that soft prompts can be obtained privately through gradient descent on downstream data. However, this is not the case for discrete prompts. Thus, we orchestrate a noisy vote among an ensemble of LLMs presented with different prompts, i.e., a flock of stochastic parrots. The vote privately transfers the flock's knowledge into a single public prompt. We show that LLMs prompted with our private algorithms closely match the non-private baselines. For example, using GPT3 as the base model, we achieve a downstream accuracy of 92.7% on the sst2 dataset with ($\epsilon=0.147, \delta=10^{-6}$)-differential privacy vs. 95.2% for the non-private baseline. Through our experiments, we also show that our prompt-based approach is easily deployed with existing commercial APIs.
Federated learning (FL) provides a variety of privacy advantages by allowing clients to collaboratively train a model without sharing their private data. However, recent studies have shown that private information can still be leaked through shared gradients. To further minimize the risk of privacy leakage, existing defenses usually require clients to locally modify their gradients (e.g., differential privacy) prior to sharing with the server. While these approaches are effective in certain cases, they regard the entire data as a single entity to protect, which usually comes at a large cost in model utility. In this paper, we seek to reconcile utility and privacy in FL by proposing a user-configurable privacy defense, RecUP-FL, that can better focus on the user-specified sensitive attributes while obtaining significant improvements in utility over traditional defenses. Moreover, we observe that existing inference attacks often rely on a machine learning model to extract the private information (e.g., attributes). We thus formulate such a privacy defense as an adversarial learning problem, where RecUP-FL generates slight perturbations that can be added to the gradients before sharing to fool adversary models. To improve the transferability to un-queryable black-box adversary models, inspired by the idea of meta-learning, RecUP-FL forms a model zoo containing a set of substitute models and iteratively alternates between simulations of the white-box and the black-box adversarial attack scenarios to generate perturbations. Extensive experiments on four datasets under various adversarial settings (both attribute inference attack and data reconstruction attack) show that RecUP-FL can meet user-specified privacy constraints over the sensitive attributes while significantly improving the model utility compared with state-of-the-art privacy defenses.
When an adversary provides poison samples to a machine learning model, privacy leakage, such as membership inference attacks that infer whether a sample was included in the training of the model, becomes effective by moving the sample to an outlier. However, the attacks can be detected because inference accuracy deteriorates due to poison samples. In this paper, we discuss a \textit{backdoor-assisted membership inference attack}, a novel membership inference attack based on backdoors that return the adversary's expected output for a triggered sample. We found three crucial insights through experiments with an academic benchmark dataset. We first demonstrate that the backdoor-assisted membership inference attack is unsuccessful. Second, when we analyzed loss distributions to understand the reason for the unsuccessful results, we found that backdoors cannot separate loss distributions of training and non-training samples. In other words, backdoors cannot affect the distribution of clean samples. Third, we also show that poison and triggered samples activate neurons of different distributions. Specifically, backdoors make any clean sample an inlier, contrary to poisoning samples. As a result, we confirm that backdoors cannot assist membership inference.
Neural network-based image classifiers are powerful tools for computer vision tasks, but they inadvertently reveal sensitive attribute information about their classes, raising concerns about their privacy. To investigate this privacy leakage, we introduce the first Class Attribute Inference Attack (CAIA), which leverages recent advances in text-to-image synthesis to infer sensitive attributes of individual classes in a black-box setting, while remaining competitive with related white-box attacks. Our extensive experiments in the face recognition domain show that CAIA can accurately infer undisclosed sensitive attributes, such as an individual's hair color, gender, and racial appearance, which are not part of the training labels. Interestingly, we demonstrate that adversarial robust models are even more vulnerable to such privacy leakage than standard models, indicating that a trade-off between robustness and privacy exists.
To efficiently monitor the growth and evolution of a particular wildlife population, one of the main fundamental challenges to address in animal ecology is the re-identification of individuals that have been previously encountered but also the discrimination between known and unknown individuals (the so-called "open-set problem"), which is the first step to realize before re-identification. In particular, in this work, we are interested in the discrimination within digital photos of beluga whales, which are known to be among the most challenging marine species to discriminate due to their lack of distinctive features. To tackle this problem, we propose a novel approach based on the use of Membership Inference Attacks (MIAs), which are normally used to assess the privacy risks associated with releasing a particular machine learning model. More precisely, we demonstrate that the problem of discriminating between known and unknown individuals can be solved efficiently using state-of-the-art approaches for MIAs. Extensive experiments on three benchmark datasets related to whales, two different neural network architectures, and three MIA clearly demonstrate the performance of the approach. In addition, we have also designed a novel MIA strategy that we coined as ensemble MIA, which combines the outputs of different MIAs to increase the attack accuracy while diminishing the false positive rate. Overall, one of our main objectives is also to show that the research on privacy attacks can also be leveraged "for good" by helping to address practical challenges encountered in animal ecology.
Federated learning (FL) was originally regarded as a framework for collaborative learning among clients with data privacy protection through a coordinating server. In this paper, we propose a new active membership inference (AMI) attack carried out by a dishonest server in FL. In AMI attacks, the server crafts and embeds malicious parameters into global models to effectively infer whether a target data sample is included in a client's private training data or not. By exploiting the correlation among data features through a non-linear decision boundary, AMI attacks with a certified guarantee of success can achieve severely high success rates under rigorous local differential privacy (LDP) protection; thereby exposing clients' training data to significant privacy risk. Theoretical and experimental results on several benchmark datasets show that adding sufficient privacy-preserving noise to prevent our attack would significantly damage FL's model utility.
Previous work has shown that Large Language Models are susceptible to so-called data extraction attacks. This allows an attacker to extract a sample that was contained in the training data, which has massive privacy implications. The construction of data extraction attacks is challenging, current attacks are quite inefficient, and there exists a significant gap in the extraction capabilities of untargeted attacks and memorization. Thus, targeted attacks are proposed, which identify if a given sample from the training data, is extractable from a model. In this work, we apply a targeted data extraction attack to the SATML2023 Language Model Training Data Extraction Challenge. We apply a two-step approach. In the first step, we maximise the recall of the model and are able to extract the suffix for 69% of the samples. In the second step, we use a classifier-based Membership Inference Attack on the generations. Our AutoSklearn classifier achieves a precision of 0.841. The full approach reaches a score of 0.405 recall at a 10% false positive rate, which is an improvement of 34% over the baseline of 0.301.
A novel form of inference attack in vertical federated learning (VFL) is proposed, where two parties collaborate in training a machine learning (ML) model. Logistic regression is considered for the VFL model. One party, referred to as the active party, possesses the ground truth labels of the samples in the training phase, while the other, referred to as the passive party, only shares a separate set of features corresponding to these samples. It is shown that the active party can carry out inference attacks on both training and prediction phase samples by acquiring an ML model independently trained on the training samples available to them. This type of inference attack does not require the active party to be aware of the score of a specific sample, hence it is referred to as an agnostic inference attack. It is shown that utilizing the observed confidence scores during the prediction phase, before the time of the attack, can improve the performance of the active party's autonomous model, and thus improve the quality of the agnostic inference attack. As a countermeasure, privacy-preserving schemes (PPSs) are proposed. While the proposed schemes preserve the utility of the VFL model, they systematically distort the VFL parameters corresponding to the passive party's features. The level of the distortion imposed on the passive party's parameters is adjustable, giving rise to a trade-off between privacy of the passive party and interpretabiliy of the VFL outcomes by the active party. The distortion level of the passive party's parameters could be chosen carefully according to the privacy and interpretabiliy concerns of the passive and active parties, respectively, with the hope of keeping both parties (partially) satisfied. Finally, experimental results demonstrate the effectiveness of the proposed attack and the PPSs.
Although recent network representation learning (NRL) works in text-attributed networks demonstrated superior performance for various graph inference tasks, learning network representations could always raise privacy concerns when nodes represent people or human-related variables. Moreover, standard NRLs that leverage structural information from a graph proceed by first encoding pairwise relationships into learned representations and then analysing its properties. This approach is fundamentally misaligned with problems where the relationships involve multiple points, and topological structure must be encoded beyond pairwise interactions. Fortunately, the machinery of topological data analysis (TDA) and, in particular, simplicial neural networks (SNNs) offer a mathematically rigorous framework to learn higher-order interactions between nodes. It is critical to investigate if the representation outputs from SNNs are more vulnerable compared to regular representation outputs from graph neural networks (GNNs) via pairwise interactions. In my dissertation, I will first study learning the representations with text attributes for simplicial complexes (RT4SC) via SNNs. Then, I will conduct research on two potential attacks on the representation outputs from SNNs: (1) membership inference attack, which infers whether a certain node of a graph is inside the training data of the GNN model; and (2) graph reconstruction attacks, which infer the confidential edges of a text-attributed network. Finally, I will study a privacy-preserving deterministic differentially private alternating direction method of multiplier to learn secure representation outputs from SNNs that capture multi-scale relationships and facilitate the passage from local structure to global invariant features on text-attributed networks.
Diffusion models have attracted attention in recent years as innovative generative models. In this paper, we investigate whether a diffusion model is resistant to a membership inference attack, which evaluates the privacy leakage of a machine learning model. We primarily discuss the diffusion model from the standpoints of comparison with a generative adversarial network (GAN) as conventional models and hyperparameters unique to the diffusion model, i.e., time steps, sampling steps, and sampling variances. We conduct extensive experiments with DDIM as a diffusion model and DCGAN as a GAN on the CelebA and CIFAR-10 datasets in both white-box and black-box settings and then confirm if the diffusion model is comparably resistant to a membership inference attack as GAN. Next, we demonstrate that the impact of time steps is significant and intermediate steps in a noise schedule are the most vulnerable to the attack. We also found two key insights through further analysis. First, we identify that DDIM is vulnerable to the attack for small sample sizes instead of achieving a lower FID. Second, sampling steps in hyperparameters are important for resistance to the attack, whereas the impact of sampling variances is quite limited.
Explainable Artificial Intelligence (XAI) aims to uncover the decision-making processes of AI models. However, the data used for such explanations can pose security and privacy risks. Existing literature identifies attacks on machine learning models, including membership inference, model inversion, and model extraction attacks. These attacks target either the model or the training data, depending on the settings and parties involved. XAI tools can increase the vulnerability of model extraction attacks, which is a concern when model owners prefer black-box access, thereby keeping model parameters and architecture private. To exploit this risk, we propose AUTOLYCUS, a novel retraining (learning) based model extraction attack framework against interpretable models under black-box settings. As XAI tools, we exploit Local Interpretable Model-Agnostic Explanations (LIME) and Shapley values (SHAP) to infer decision boundaries and create surrogate models that replicate the functionality of the target model. LIME and SHAP are mainly chosen for their realistic yet information-rich explanations, coupled with their extensive adoption, simplicity, and usability. We evaluate AUTOLYCUS on six machine learning datasets, measuring the accuracy and similarity of the surrogate model to the target model. The results show that AUTOLYCUS is highly effective, requiring significantly fewer queries compared to state-of-the-art attacks, while maintaining comparable accuracy and similarity. We validate its performance and transferability on multiple interpretable ML models, including decision trees, logistic regression, naive bayes, and k-nearest neighbor. Additionally, we show the resilience of AUTOLYCUS against proposed countermeasures.
The privacy preserving properties of Langevin dynamics with additive isotropic noise have been extensively studied. However, the isotropic noise assumption is very restrictive: (a) when adding noise to existing learning algorithms to preserve privacy and maintain the best possible accuracy one should take into account the relative magnitude of the outputs and their correlations; (b) popular algorithms such as stochastic gradient descent (and their continuous time limits) appear to possess anisotropic covariance properties. To study the privacy risks for the anisotropic noise case, one requires general results on the relative entropy between the laws of two Stochastic Differential Equations with different drifts and diffusion coefficients. Our main contribution is to establish such a bound using stability estimates for solutions to the Fokker-Planck equations via functional inequalities. With additional assumptions, the relative entropy bound implies an $(\epsilon,\delta)$-differential privacy bound or translates to bounds on the membership inference attack success and we show how anisotropic noise can lead to better privacy-accuracy trade-offs. Finally, the benefits of anisotropic noise are illustrated using numerical results in quadratic loss and neural network setups.