Boundary-targeted Membership Inference Attacks on Safety Classifiers
Authors: Anthony Hughes, Alexander Goldberg, Prince Jha, Adam Perer, Nikolaos Aletras, Niloofar Mireshghallah
Organizations: University of Sheffield · Carnegie Mellon University · MBZUAI
Abstract
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.
Membership inference attacks (MIAs) are a canonical way to assess a machine learning model's privacy properties. Although several attempts have been made to evaluate MIAs on language models, the extant literature has suffered numerous difficulties in constructing clean evaluations to test new techniques. In particular, subtle distribution shifts between member and non-member sets can undermine the statistical validity of MIAs; recent work has underscored this by showing that "blind" methods with no access to the underlying model can perform far better than published methods on the same benchmarks. This paper constructs a benchmark for principled evaluation of MIAs against LLMs, by leveraging the insight that training data before and after a fixed point during training are drawn from the same distribution. Therefore, all open-source models with intermediate checkpoints and public training data can be converted into MIA testbeds. We apply our framework to a half-dozen published attacks on the Pythia and OLMo family of models, from 70M to 7B parameters. To facilitate further privacy research, we open-source a modular library for designing and implementing attacks in this setting: https://github.com/safr-ai-lab/pandora_llm.
Membership inference attacks (MIAs) try to determine whether a specific record was used to train a model, a privacy risk that matters in natural language processing (NLP), where training data can contain sensitive user text. This paper presents a controlled benchmark of membership inference vulnerability for text classification on the GLUE SST-2 sentiment dataset. A TF-IDF + Logistic Regression pipeline and a fine-tuned DistilBERT classifier are compared under a loss-threshold MIA, with utility measured by development accuracy and macro F1. DistilBERT reached 0.9466 accuracy and 0.9460 macro F1 against 0.8756 and 0.8727 for Logistic Regression, yet both models leaked membership signal (Attack AUC 0.5615 and 0.5800, respectively). Two mitigations were tested. Stronger regularization reduced leakage for Logistic Regression at a visible utility cost, whereas fine-tuning DistilBERT for 2 epochs instead of 3 reduced leakage with negligible accuracy loss. Lightweight training adjustments can improve the privacy-utility trade-off without complex defenses.
Large generative models across text-to-text, text-to-image, and image-to-text modalities have been shown to pose significant privacy risks. One fundamental threat is membership inference attacks (MIA), which aim to determine whether a given data point was used in a model's training set. Although prior work has investigated MIAs against these three classes of generative models, existing approaches treat them in isolation and are not cross-applicable, thereby limiting their real-world utility. To address this limitation, we present the first comprehensive study of a unified membership inference framework that applies across text-to-text, text-to-image, and image-to-text modalities. Our approach is grounded in a key modality-agnostic observation: the output distribution of a generative model can approximate its training data distribution. Leveraging this property, we model the distributions of model-generated outputs and auxiliary non-member samples in a shared embedding space, and perform membership inference via likelihood ratio testing. We conduct extensive experiments in a strict black-box setting under both partial-knowledge and zero-knowledge threat models, and evaluate membership inference against both fine-tuning and pre-training data. Experimental results demonstrate our approach's superior performance in comparison to existing state-of-the-art methods, which are typically optimized for a single model class.