Vision Transformers (ViTs) are increasingly used in split-inference systems, where edge devices transmit intermediate token representations to a remote cloud. In this setting, token reduction lowers computation and communication costs, while token shuffling disrupts the spatial organization of the transmitted tokens, potentially limiting information leakage. However, their privacy benefits remain unclear against feature inversion attacks, which attempt to reconstruct the input from the transmitted embeddings. In this work, we show that, despite disrupting the spatial structure required by conventional reconstruction attacks, transmitted token embeddings retain substantial positional information. Based on this observation, we introduce the Spatially Aligned Reconstruction Attack (SARA), a unified pipeline that predicts token positions, restores their spatial layout, reconstructs missing embeddings using a feature-space masked autoencoder, and recovers the input image. Our results demonstrate that token shuffling provides only apparent privacy, as SARA largely reconstructs the original token organization. Token reduction offers stronger protection, but significant leakage persists when the retained tokens preserve sufficient semantic and positional information. Finally, we introduce a lightweight edge-side defense that removes positional embeddings and progressively adapts the edge-side transformer blocks through knowledge distillation. It substantially reduces attack performance against SARA, while preserving downstream task accuracy and requiring no changes to the cloud-side model.
Large language models (LLMs) have recently advanced graph neural networks (GNNs) by enriching node representations with semantic information, giving rise to LLM-enhanced GNNs that achieve substantial performance gains. However, their vulnerability to privacy attacks, in which adversaries infer sensitive information from model outputs, remains largely underexplored. To bridge this gap, we present a systematic evaluation of privacy risks in LLM-enhanced GNNs through a unified framework consisting of five stages: (1) dataset preparation, (2) victim model training, (3) privacy attack, (4) risk assessment, and (5) defense analysis. Specifically, we conduct experiments on six real-world text-attributed graph datasets covering diverse domains. We consider six representative privacy attack methods targeting three fundamental threats, namely link, label, and membership inference, and construct 42 victim model configurations by combining multiple LLM-based feature enhancers with representative GNN backbones. Extensive experiments show that, despite their utility improvements, LLM-enhanced GNNs consistently exhibit increased vulnerability to privacy attacks compared to shallow text representation baselines. Further analysis reveals that semantic enrichment amplifies link-, label-, and membership-related signals in the embedding space, making them more exploitable by inference attacks. Finally, we evaluate differential privacy as a defense strategy and show that, while it can partially mitigate privacy risks, it introduces significant utility degradation, highlighting a fundamental privacy-utility trade-off in LLM-enhanced graph learning. Overall, this work provides a comprehensive understanding of privacy risks in LLM-enhanced GNNs and offers practical insights for developing more secure and trustworthy graph learning systems.
When a person's records appear in k independent data silos, each protected by (epsilon, delta)-differential privacy, standard composition yields a valid (k*epsilon, k*delta)-DP guarantee for the joint output. This worst-case bound, however, does not answer the concrete inference question: at what k can an adversary actually identify a target person? This paper develops the information-theoretic framework needed to answer that question. We introduce cross-silo person-level DP (XSP-DP), a Pufferfish-style privacy notion whose adjacency relation captures all records of a single person across all silos simultaneously, and verify that the standard basic composition bound carries over to this adjacency model. Within this framework we prove that de-anonymization undergoes a phase transition at k* = Theta(log n / epsilon^2) (population size n, per-silo RR parameter epsilon): a Fano lower bound shows any estimator fails for k << k*, while a matching maximum-likelihood upper bound shows the attack succeeds for k >> k*. An explicit XOR + randomized-response construction demonstrates information synergy: each silo's output is individually uninformative about the target, yet the joint mutual information is strictly positive. For non-coordinated binary randomized-response mechanisms, we prove that de-anonymization is inevitable once k exceeds the threshold, establishing that cross-silo coordination is necessary. These results provide a baseline threat model and Theta-level threshold for cross-silo inference attacks under local DP.
Counterfactuals are typically used in high-stakes decision areas to explain a machine learning model by showing how changes to the user profiles result in the desired outcome. However, explaining the model's decisions through counterfactuals can also be exploited by an adversary to conduct privacy attacks against the model or its training data. Drawing on the analogy that counterfactuals provide realistic substitutes for real training data, similar to synthetic data, we demonstrate in this paper how it is possible to successfully perform privacy attacks on counterfactuals by drawing on the attacks developed against synthetic data. More precisely, we investigate the effectiveness of the membership inference attacks designed for synthetic data on various types of counterfactuals. Additionally, while existing membership inference attacks against counterfactuals usually require to be able to query the model, we show how it is possible to perform successful membership inference attacks using only a set of counterfactuals, with no access to the model from which they are generated. Our results demonstrate that model developers should be more cautious when releasing counterfactuals to various users, as it can lead to a privacy breach.
Anthony Hughes, Alexander Goldberg, Prince Jha +3cs.LG cs.CL
Safety classifiers are essential safeguards within generative AI systems, filtering harmful content or identifying at-risk users when interacting with large language models. Despite their necessity, these models are trained on sensitive datasets including discussions of self-harm and mental health, raising important, yet poorly understood, privacy concerns. Membership inference attacks (MIAs) allow adversaries to infer membership of examples used to train models. In this work, we hypothesize that identifying the examples on which the classifier is least confident are informative for an adversary to infer membership. This reflects a localized failure of generalization, where the model relies on memorization to resolve ambiguity in the training set. To investigate this, we introduce a new boundary-targeted selection strategy that identifies low confidence examples that amplify the signal of an examples membership within a training set. Our experimental results show that an adversary can recover 19\% of the conversations a safety classifier flagged as indicating user distress, at a 5\% false-positive rate, on a classifier fine-tuned for detecting a user who may require emotional support. This is $3.5$ times more than attacking using state-of-the-art MIA methods alone. Finally, we characterize the boundary laying examples and show that content-based filtering is ineffective for protection, and existing noise strategies can effectively mitigate susceptibility of these examples.
Large language models (LLMs) are increasingly deployed in interactive and retrieval-augmented settings, raising significant privacy concerns. While attacks such as Membership Inference (MIA), Attribute Inference (AIA), Data Extraction (DEA), and Backdoor Attacks (BA) have been studied, they are typically analyzed in isolation, leaving a gap in understanding their behavior under common system factors. In this paper, we introduce a unified threat model and notation, reproduce a representative set of privacy attacks, and conduct a structured ablation study to evaluate the impact of key factors such as model architecture, scale, dataset characteristics, and retrieval configuration. Our analysis reveals clear differences across attack types. Membership inference attacks, particularly mask-based variants, exhibit strong and reliable signals, while backdoor attacks achieve consistently high success rates due to their trigger-based nature. In contrast, attribute inference and data extraction attacks remain more challenging, resulting in lower accuracy, yet they pose significant risks as they target sensitive personal information. Overall, these results highlight that privacy risks in LLM systems are highly context-dependent and driven by design choices, emphasizing the need for holistic evaluation and informed deployment practices.