Membership Inference Attacks

Also known as MIA

Overview

Membership inference attacks ask whether a particular record, person, or dataset contributed to a model’s training, helping assess privacy risks. Black-box attacks use outputs such as loss, confidence, or generated text, while white-box methods analyze gradients and internal representations; LiRA and RMIA are prominent reference-model approaches. Researchers apply these methods to language, vision, generative, recommender, and federated models, and study defenses including differential privacy, regularization, and limits on exposed signals. A central challenge is that results can change sharply with attacker access, distribution shift, and false-positive rates, so reliable audits must distinguish practical privacy risks from apparent attack success under unrealistic assumptions.

Latest papers

125 papers14 papers in the last four weeks, up 17% on the four weeks before. 0.1% of all new papers.

Nov 13, 2026

Quantifying Membership Inference Risks in Genomic Data Sharing: The Role of Attacker's Auxiliary Information

Genomic data sharing is essential for biomedical research but introduces significant privacy risks, including membership inference attacks, in which an adversary determines whether a specific individual's data was used in a study. Existing privacy protections are typically applied after a release format has already been chosen, leaving researchers without a principled method to assess dataset-specific risk beforehand. We propose a server-assisted, principal component analysis-based pre-release auditing framework that maps genomic data into a shared low-dimensional space, applies epsilon-local differential privacy via the Laplace mechanism with a formally derived sensitivity for principal component analysis scores, and estimates membership inference attack power at a controlled 5% false-positive rate using a distance-threshold attack. A membership-neutrality diagnostic based on Unbiased Reconstruction Robustness is reported alongside attack power, precision, and recall. Experiments using the openSNP and 1000 Genomes Project datasets across multiple dimensionalities and privacy budgets show that separation between the researcher's dataset and public auxiliary data in principal component analysis space is a dominant determinant of attack success, exceeding the effect of perturbation level alone. Comparison against a genome-wide association study-style likelihood ratio test baseline further contextualizes the measured risk. This framework enables researchers, data repositories, and institutional review bodies to quantify dataset-specific membership inference risk before genomic data publication, supporting more informed and privacy-aware biomedical data-sharing decisions.
Nov 9, 2026

XposedMIA: Exploiting Explainable Artificial Intelligence (XAI) for One-Shot Membership Inference Attacks

Machine learning (ML) services increasingly expose explainability outputs alongside model predictions. While such explanations are intended to improve transparency, they may also reveal information about the underlying training data. In this work, we show that the local surrogate geometry employed by Local Interpretable Model-agnostic Explanations (LIME) can be repurposed as a proxy for a sample's distance to the model's decision boundary. We extend this boundary-distance view to tabular counterfactual explanations and, for the first time, to visual counterfactual explanations for image classifiers. We present XposedMIA, a one-shot, explanation-based black-box membership inference attack (MIA) framework that converts the returned explanation into a proxy for the queried sample's distance to the model decision boundary. After a one-time calibration step on a small labeled calibration set of as few as 30 samples, our attack requires only a single query for each target sample. Unlike prior label-only attacks requiring iterative queries or shadow training, our method relies solely on the predicted label and its accompanying explanation. Our evaluation indicates that XposedMIA achieves high effectiveness, with success rates of up to 0.95 using only a single query per target, revealing that releasing explanations introduces an additional channel for membership leakage.
Nov 8, 2026

A Theoretical Reduction From Membership Inference to Reconstruction Attacks Against Learners

We study reconstruction attacks, which recover training data from a learned model. In our adversarial model, the adversary designs a hard instance, drawn from a parameterized family of data distributions, and is given the parameter used to generate it. Our main contribution is a general reduction showing that, in this model, membership inference (determining whether a given datapoint was used for training) implies reconstruction: for a training set of size nn, a membership test with recall at least q(n)q(n) and false-positive rate O(δ2/n2)O(δ^2/n^2) yields a reconstruction algorithm that, with probability 1−O(δ)1 - O(δ), outputs only training points, and recovers Ω(n⋅q(n))Ω(n \cdot q(n)) of them in expectation. This reduction applies to any learning algorithm. For stochastic convex optimization (SCO), we make the reduction computationally efficient over exponentially large domains via structured distributions whose support is partitioned into polynomially many categories of exponentially many elements each, so that reconstruction iterates over categories rather than the full support. Combining this with an existing membership inference result for SCO, we prove that any learning algorithm achieving low excess risk over all distributions admits a polynomial-time reconstruction attack recovering Ω(n)Ω(n) training points, even when the underlying distribution has exponentially large support.
Nov 7, 2026

Bayesian Low-Rank Posteriors for Scalable Membership Inference

Membership inference attacks (MIAs) aim to determine whether a sample was used during the training of a target model. Strong MIAs rely on reference (or shadow) models to estimate target-conditional score distributions, but training such ensembles is computationally expensive for modern large-scale Vision--Language Models (VLMs). In this work, we propose a scalable alternative that replaces explicit reference-model training with a Bayesian approximation over low-rank adaptation parameters. Specifically, we model uncertainty only within the LoRA subspace and construct a posterior from a single training run, from which we sample a diverse set of virtual reference models. We instantiate this approach using Stochastic Weight Averaging Gaussian (SWAG), enabling efficient approximation of target-conditional score distributions at a fraction of the cost of conventional shadow-model ensembles. We evaluate the resulting attack on three downstream-adapted VLMs across two privacy-sensitive visual question answering tasks, namely Document VQA and Medical VQA. To isolate the effect of pretraining knowledge, we additionally introduce a controlled synthetic dataset. Our approach outperforms standard reference-based attacks while requiring only a single reference-model fine-tuning run, making membership inference more practical for large VLMs.
Nov 6, 2026

Overview of the ImageCLEFmedical 2026 GANs task: Exploring fingerprint detection, privacy leakage, and synthetic medical image generation

The 2026 ImageCLEFmedical GANs Task Exploring Fingerprint Detection, Privacy Leakage, and Synthetic Medical Image Generation continues to investigate privacy and security concerns around using patient data to generate synthetic medical images, now in its fourth edition. This edition introduces three complementary subtasks: the first extends prior editions by asking participants to detect which real images were used in training a Generative Adversarial Network to produce given synthetic outputs; the second is a novel subtask requiring participants to identify which latent vector produced each synthetic image, probing the faithfulness and reversibility of the latent-to-image mapping; the third asks participants to generate privacy-preserving CT slices, quantifying the trade-off between image fidelity and patient privacy. Ground-truth annotations and benchmark datasets of real and GAN-generated lung CT slices are provided for all subtasks. Evaluation is based on Cohen's Kappa for Subtask 1, matching accuracy for Subtask 2, and a combination of FID and Privacy Preservation Score (PPS) for Subtask 3. Overall, 37 teams registered for the task, with 10 teams submitting runs across subtasks and 6 teams submitting working notes papers. This paper presents an overview of the task setup, datasets, evaluation metrics, and summarises the approaches and results of participating teams.
Nov 4, 2026

Timing is Everything: Inferring Temporal Membership in Continually Updated Models

Continual model updating has become a common practice in the deployment of machine learning systems. However, the dynamic training behaviors give rise to new risks not sufficiently addressed by existing studies. In this paper, we present the first systematic study of temporal privacy risk in continually updated deep learning models. We formalize the notion of temporal membership in continually updated models. We then define the temporal membership inference attack, which seeks to infer whether and when individual samples were involved in model training, at both timestamp-level and interval-level granularities. To this end, we propose TMIA, a novel black-box attack that leverages bidirectional LSTMs to infer temporal membership from model outputs observed across successive updates. Through extensive experiments on three language models and three image classification models, we show that TMIA is highly effective and significantly outperforms existing baselines. Finally, we evaluate three potential defenses, revealing the trade-off between privacy protection and model utility.
Nov 2, 2026

Probing Local Conditional Fit for Diffusion Model Membership Inference

Membership inference determines whether a candidate example was used to train a text-to-image diffusion model and is an important tool for auditing training-data use. We identify a directional membership signal in conditional prediction shifts, the changes in predicted noise under alternative text conditions at a fixed noisy input: whether movement toward the alternative prediction initially lowers or raises the denoising loss from its value under the original caption. Members exhibit less room for such local improvement. To quantify this response, we propose LoFIT (Local Conditional Fit), which measures the initial response of the denoising loss along normalized prediction shifts and aggregates it across multiple text conditions and diffusion states. LoFIT requires no target-model gradients and can be computed using only forward queries to the denoiser. Evaluations on four datasets show improved membership discrimination across fine-tuned and pretrained models. LoFIT distinguishes members early in fine-tuning, while the naive denoising-loss baseline remains near chance. Controlled comparisons show that the membership signal is concentrated in the first-order response, with little discrimination from prediction-shift magnitude alone.
Oct 29, 2026

Characterizing Recoverable Expert-Utilization Signatures in Fine-Tuned Mixture-of-Experts LLMs

As Mixture-of-Experts (MoE) large language models are increasingly deployed in distributed environments, from federated learning to edge-cloud systems, their architecture harbors an underexplored landscape of security vulnerabilities. Moving beyond textual analysis, we uncover structural memorization: a novel vulnerability where a fine-tuned model memorizes not only its training data but also the internal pathway used to process it, creating a characteristic routing fingerprint unique to the training process. We provide comprehensive empirical evidence that structural memorization emerges consistently even from a few samples, and that its intensity is directly influenced by the choice of trainable modules. We show that these memorized expert pathways can be stealthily recovered through the model's unconditional generation, using only routing and output logits, with no knowledge of the fine-tuning data or procedure. In distributed deployments, this vulnerability is amplified as routing information is exposed at multiple trust boundaries, including federated aggregation points, edge-cloud interfaces, and multi-tenant monitoring systems. Recovery exposes a critical blind spot: traditional membership inference metrics, viewed through the lens of expert routing, fail to identify memorized data corresponding to these structural pathways. To close this gap, we devise composite metrics combining reference model tail likelihood with complexity-aware normalization, significantly improving detection of memorized routing structures. This enables adversaries to pinpoint critical experts shaped by fine-tuning, whether dominant or dormant, and execute targeted attacks such as router manipulation or backdoor injection in realistic distributed settings where attackers lack training data but can observe routing patterns across system components.
Oct 29, 2026

Difficulty Calibration Across Time for Membership Inference in Diffusion Models

Modern diffusion models are trained on massive datasets that often include copyrighted or private images. Membership inference attacks (MIAs) address this by testing whether a given image was in a model's training set. The standard approach scores an image by its reconstruction loss and flags low-loss images as members. This often fails because some images are simply easier to reconstruct than others, so a simple unseen image can score a lower loss than a complex training image. Difficulty calibration separates memorization from how difficult an image is to begin with. We show how to calibrate along two axes of time: training time and diffusion time. First, across training time, we introduce Reference-Model Calibration (RMC), which compares a fine-tuned model's loss against an earlier checkpoint. This cleanly isolates newly memorized data and gives near-perfect auditing, even without text captions. Since earlier checkpoints are not always available, we next present Temporal Self-Calibration (TSC), which calibrates across diffusion time using only a single model. Our key observation is that an image's difficulty persists across noise levels, whereas memorization is localized in time. By comparing the loss at a target timestep against neighboring timesteps from the same model, TSC attenuates image difficulty using the model's own trajectory. Empirically, our methods beat the state of the art by a wide margin, and unlike most existing attacks, generalize across different model parameterizations such as xx-, vv-, and εε-prediction models. Code is available at https://github.com/MichaelJoseph777/TSC.
Oct 27, 2026

A Simple Presentation of Robust Membership Inference Attacks

In a membership inference attack, an attacker is given the data of a particular individual, and tries to determine if that individual was included in a particular dataset, given some kind of statistics derived from the dataset or a model trained on the dataset. Roughly, a membership inference attack is robust if it succeeds even when the statistics are highly noisy, regardless of how that noise is introduced. We give a simple presentation of robust membership inference attacks for mean estimation, covering both multivariate Gaussian distributions and binary product distributions. We also give a summary of the literature on robust membership inference, covering history, extensions, and connections to other research areas.
Oct 22, 2026

OP-Print: Tracing Membership Imprints in On-Policy Self-Distillation

On-policy distillation has been widely adopted for alleviating the distribution mismatch of off-policy distillation. Self-distillation further removes the need for an external teacher by using the student's own outputs, enhanced with extra context or demonstrations, as supervision. This teacher-free design makes on-policy self-distillation practical, yet its privacy risks remain largely unexplored. In particular, using two coupled views of the same training example may strengthen its memorization. Motivated by this, we present the first study of membership inference attacks against on-policy self-distillation. We investigate both gray-box and black-box threat models and propose a unified attack framework, named OP-Print. OP-Print builds on our finding that the supervision imprint in on-policy self-distillation is sparse and structured, concentrating on uncertainty-sensitive correction sites rather than across the entire rollout. Exploiting this property, we construct attack signals tailored to the two access levels: in the gray-box setting, OP-Print directly aggregates the privileged teacher-student residual correction along the same rollout; in the black-box setting, it measures restart-based output-basin stability as an observable surrogate for the hidden correction process. Across diverse models and distillation schemes, all baseline attacks remain essentially at random guessing even under gray-box access, whereas OP-Print is broadly validated to be effective in both gray-box and black-box settings.
Oct 22, 2026

Shift-aware score transport for privacy auditing without retraining

Privacy auditing empirically assesses privacy leakage by using membership inference attacks to establish lower bounds on privacy loss. We consider an observational auditing setting in which the auditor cannot insert canaries, control data inclusion, or retrain the target model, and must instead rely on fixed sets of known members and non-members. This setting is relevant for large deployed models or when retraining is restricted by computational or access constraints. In this case, non-members available for auditing may differ in distribution from the training data, causing membership-inference scores to be skewed due to the distribution shift. We address this problem by estimating the counterfactual in-distribution non-member score distribution using the available out-of-distribution (OOD) non-members. Our proposed method learns the relationship between data representations and membership-inference scores from OOD non-members and transports it to the target covariate distribution. Because this transport may become unreliable under extrapolation or imperfect conditional invariance, we further introduce a conservative correction calibrated directly to the nominal transported privacy level. We construct controlled calibration tasks from the available non-member data to estimate how representation-space displacement inflates the audit, and use this estimate to correct the transport-induced audit error. Across multiple datasets with distribution shifts, our method yields substantially more stable privacy estimates while retaining tighter lower bounds than baselines.
Oct 21, 2026

Automated screening of privacy risks in medical AI

Medical artificial intelligence (AI) models are often shared without their training data, assuming this protects the patients who contributed them. However, a model can reveal whether a patient's records were used in training, a privacy breach known as a membership inference attack. This risk is usually reported as a cohort average, which can appear negligible while individual patients are almost certainly exposed. Identifying potentially vulnerable records before developing a production model could make privacy assessment more practical. We measured each patient's vulnerability to membership inference attack, across 7,789 models trained on six clinical datasets in four modalities, including two adapted pretrained medical models. Training choices changed how many patients had high attack scores, but the patients who ranked highest by risk largely stayed the same across the configurations we tested. Stronger weight decay reduced the exposed cohort up to threefold, yet largely the same patients remained most at risk across training recipes, architectures and initializations, including foundation models. Therefore, we developed a lightweight screening tool that uses four small models to identify patients with high risk of exposure before production-model training. We also assessed if flagged patients could be protected without removing their data. Privacy screening should become incorporated into data preparation for clinical AI.
Oct 19, 2026

Reliable Estimation of Differential Privacy under Selection Bias

Statistical privacy audits are essential for verifying the theoretical privacy guarantees of differentially private (DP) mechanisms. Auditors typically use Membership Inference Attacks (MIAs) to estimate a lower bound on the true privacy leakage of DP mechanisms at a specified confidence level, which should not exceed the theoretical privacy upper bound guaranteed by DP. However, we found that many auditing implementations in prior work introduce a selection bias when running MIAs, which leads to statistically invalid and overestimated lower bounds. Because most audits are not tight, this bias has thus far gone unnoticed: the estimated lower bounds remained well below the theoretical upper bounds. Nevertheless, accurately evaluating this lower bound is critical, as a flawed statistical assessment could lead to false alarms on correct mechanisms or an incorrect evaluation of the guarantee's tightness. This work employs tools from the statistical literature to correct this bias. We present an overview of the different correction methods and assess their respective advantages. Furthermore, we perform a comprehensive re-analysis of results returned by prior audit implementations and show that without statistical correction, these audits violate the theoretical upper bounds up to four times more often than their claimed confidence level implies. With correction applied, the resulting lower bounds are statistically valid, and violations occur at no more than the nominal rate.
Oct 17, 2026

Frequency-Domain Membership Inference Does Not Beat Confidence Thresholding: A Band-Resolved Analysis of Why

A widely repeated account of membership inference holds that models lean on fine-grained detail for samples they have memorized, so degrading an input's resolution should expose membership through how quickly a prediction falls apart. We test that account band by band. A monotone family of zero-phase low-pass filters is applied to each query image, the target's confidence on its original prediction is read at every step, and the resulting decay is scored in a dozen ways. Two federated ResNet-18 targets are studied, one trained normally and one pushed to a 45.1-point generalization gap. The account does not survive. On the normally trained target the best profile score reaches an AUC of 0.513 against 0.537 for a single confidence threshold, and on the overfit target 0.735 against 0.943. A search over 1206 candidate scores, with direction chosen on a held-out half and a fitted combination as an upper bound, produces no score that beats the threshold in either regime. The failure has a mechanism: the mean confidence gap between members and non-members grows as detail is removed, exactly as the hypothesis predicts, but the spread grows faster, so discriminability falls from the first step onward. Decomposing each image into phase and magnitude shows why no band-localized operation can work, since the advantage travels with image structure, surviving at 0.567 when the phase is kept and collapsing to 0.492 when only the magnitude is. Membership is a whole-image identity effect rather than an accumulation of spectral energy in any band, and defenses aimed at high-frequency reliance are aimed at the wrong target.
Oct 13, 2026

SoK: Rethinking Membership Inference in Tabular Machine Learning

Membership Inference Attacks (MIAs) are currently the dominant approach to evaluating privacy in machine learning applications. Despite their significance, several concerns remain unexplored, particularly with regard to tabular data. Existing evaluations focus predominantly on high-dimensional datasets, leaving the privacy behavior of small-to-medium datasets unexplored. In this paper, we present a new taxonomy for MIAs and conduct a comprehensive study of the factors influencing vulnerability. We evaluate four distinct state-of-the-art attacks across diverse tabular datasets and model architectures. Our results show that datasets with higher class granularity and majority class percentage are more susceptible to MIAs. We further demonstrate that advanced MIAs are highly dependent on adversarial knowledge of the target model, while reference model-free attacks often achieve statistically comparable results. Most importantly, we show that records with unique signatures (single-outs) remain highly vulnerable, even when common defenses are applied. These findings indicate that current defenses reduce average leakage, but fail to adequately protect the most sensitive individuals.
Oct 12, 2026

COMPASS: Membership Inference Attack Against Vision Large Language Models via Adaptive Noise Calibration

Vision large language models (VLLMs) are deployed in privacy-sensitive domains, raising training-data exposure concerns. Existing gray-box membership inference attacks (MIAs) commonly rely on entropy statistics or generated descriptions, requiring costly generation with limited membership separation. Perturbation-based methods instead measure response changes but focus on large perturbations, leaving VLLM membership behavior under smaller perturbations poorly understood. We systematically characterize VLLM membership behavior under controlled Gaussian perturbations across noise levels at fine resolution. We find that optimal perturbation magnitude and membership-signal direction vary across model-dataset pairs. In many settings, a previously unstudied low-noise regime yields stronger separation than large perturbations, with members exhibiting lower divergence than non-members, reversing the conventional score direction. Other settings retain the conventional direction, demonstrating that neither a fixed perturbation magnitude nor a predetermined score orientation is universally reliable. We propose COMPASS, an adaptive gray-box MIA that calibrates both the perturbation magnitude and membership-score direction using a loosely distribution-matched, known non-member reference set. We instantiate COMPASS using Renyi-normalized KL divergence and Renyi divergence, with max-k aggregation over the most informative image-logit positions. Across LLaVA v1.5-7B, MiniGPT-4, and HuluMed-32B, spanning pretraining, instruction-tuning, and medical-image datasets, COMPASS outperforms baselines. On general-domain data, COMPASS achieves AUCs up to 0.88. On the distribution-matched medical evaluation, it reaches 0.64 AUC, while the strongest baseline remains at chance level. These results establish full-range adaptive perturbation analysis as an effective approach to VLLM membership inference.
Oct 8, 2026

Policy Alignment: New Signals for Membership Auditing in On-Policy Distillation

On-policy distillation (OPD) trains a student model by aligning its policy with a teacher model on trajectories generated by the student model itself. Through this process, the student policy moves toward the teacher on the prompts used for distillation. However, these prompts are often private and costly, creating a need for prompt-level membership auditing. Existing methods mainly rely on likelihood-based confidence signals or student policy drift between checkpoints, but they do not capture the teacher-induced direction of the student update. In this paper, we propose Policy Alignment Membership Auditing (PAMA), a new auditing framework tailored for OPD. Our key observation is that a member prompt directly contributes to the teacher-guided policy update, while a non-member prompt only experiences indirect effects through cross-prompt generalization. Based on this directional trace, PAMA measures whether the student update moves toward reducing the teacher loss on a candidate prompt. Specifically, we introduce Teacher Alignment Gain (TAG) to estimate the teacher-aligned update direction from model outputs, and further combine it with student drift and uncertainty alignment signals for reliable membership auditing. We evaluate PAMA on six datasets and three teacher-student model families. On MATH, the primary evaluation benchmark, PAMA achieves AUC values of 0.791--0.941, improving AUC by 14.6--20.6% over state-of-the-art baselines.
Oct 7, 2026

Power Side-Channel Membership Inference Attack on Embedded Machine Learning

Membership inference attacks (MIAs) threaten the privacy of machine learning (ML) training data by determining whether a sample was used to train a target model. Existing MIAs rely on model outputs, ranging from prediction probabilities to predicted labels, an assumption that can be restrictive for on-device ML systems with limited or inaccessible outputs. However, suppressing model outputs does not eliminate the data-dependent computations that produce them, which may remain observable through physical side channels. We present PSCMIA, a power side-channel membership inference attack against embedded ML models that can infer membership directly from power traces without requiring prediction probabilities or even the predicted labels. We evaluate PSCMIA across multiple datasets (MNIST, FMNIST, CIFAR10, CINIC10), fully connected (FC) and convolutional neural network (CNN) architectures, and two embedded platforms (STM32F3, XMEGA). PSCMIA achieves ROC-AUC values of up to 0.907 on FC models. For CNN models, the ROC-AUC gap between PSCMIA and probability vector-based shadow MIA ranges from 0.006 to 0.116. Across the FC and CNN evaluations, PSCMIA outperforms label-only MIA in 11 of 16 model-dataset-hardware configurations, demonstrating that physical execution can expose membership information even when conventional model outputs are unavailable through unintended power side-channel leakage.
Oct 7, 2026

Closed-Form Noise Calibration Against Membership Inference for Random-Allocation DP-SGD

DP-SGD protects training data by adding Gaussian noise to clipped gradients. The amount of noise is usually chosen by running a numerical privacy accountant inside a search. We study DP-SGD with random allocation, where each epoch uses every record once, at a randomly chosen step. For this setting we give a one-line formula that bounds the accuracy of every membership inference attack (MIA) on the trained model. With MM steps per epoch, EE epochs and noise multiplier σσ, and with membership and non-membership equally likely a priori, the attack accuracy is at most 12+14(1+(e1/σ2−1)/M)E−1\frac12+\frac14\sqrt{(1+(e^{1/σ^2}-1)/M)^E-1}. The formula comes from the chi-square divergence between a Gaussian distribution and a Gaussian mixture that dominates random allocation. It is interpretable and gives σσ in about a microsecond. Where applicable, our formula needs at most about half the noise of the state-of-the-art closed-form bound. To measure how close the bound is, we also derive an exact expression for the attack accuracy of these two distributions and evaluate it numerically. Calibrating to this exact expression requires 13.0%13.0\% to 20.2%20.2\% less noise than the formula in our main experiments, and since it is exact, no accountant that knows only MM, EE and σσ can certify a smaller σσ. In training, the resulting σσ outperforms the formula and matches a published accountant in test accuracy. It is found in seconds and certified in minutes, whereas every search we ran with that accountant took longer or returned at least 0.62%0.62\% more noise. We show that MIAs on the trained models stay below the bound.
Oct 7, 2026

When Routing Reveals Membership: Privacy Leakage from MoE Router Telemetry

Mixture-of-Experts (MoE) language models produce routing information during inference that may be logged or exposed for monitoring, debugging, load analysis, and safety auditing. Unlike ordinary model outputs, this telemetry reveals a view of the model's internal computation, raising a privacy question: can it reveal whether an example was used to fine-tune the deployed model? We introduce a router-augmented membership inference attack that combines conventional output-side signals with aggregated routing features and applies a membership classifier learned from independently fine-tuned shadow models to the target model. Across three MoE architectures and three data domains, router telemetry consistently improves membership inference over a strong output-signal ensemble, increasing TPR at 1% FPR by 2.7--9.4 percentage points across all nine settings. The leakage persists across full fine-tuning, frozen-router training, LoRA, and instruction tuning, and remains observable with only discrete expert selections, restricted telemetry, or a single shadow model. Mechanistic analysis further shows that the leakage does not require router-specific memorization: fine-tuning introduces membership information into hidden representations, while the router exposes a projection of this signal even when its parameters are frozen. Perturbing the telemetry reduces this additional leakage only as its fidelity degrades. Our results show that router telemetry can turn an operational signal into an additional privacy surface for fine-tuned MoE models.
Oct 4, 2026

Hidden Risks of Jev: An Empirical Study of Security, Privacy, and Dual Use

Jev turns natural-language questions into typed answers and probabilities with low latency and cost, enabling applications to route requests and select tools. While this interface allows Jev to integrate naturally into application workflows as a decision layer, the security and privacy implications of this emerging use remain largely unexplored. To address this gap, we conduct the first systematic study of these implications using the official Jev API and NanoJev, a local model with controllable training data and updates, focusing on three research questions: (1) What security threats arise when Jev is deployed as an application decision layer? (2) What private information can Jev reveal despite returning constrained typed outputs? (3) How can Jev's general-purpose decision capability be used for beneficial purposes or misused? Jev's decisions depend on application state and may be influenced by user-provided inputs. We therefore adapt prompt injection and adversarial suffixes to manipulate its decisions. Open-source Jev distribution and updates introduce supply-chain risks, which we examine by implanting backdoors in NanoJev through training data poisoning. Since Jev's outputs reflect both application state and information learned during training, we further adapt membership, private attribute, and internal knowledge inference attacks to recover sensitive information despite its constrained output format. Finally, Jev can serve as a general-purpose decision oracle for defensive and malicious workflows. We examine this dual use through four detection tasks covering prompt injection, jailbreak inputs, harmful content, and AI-generated text, alongside misuse scenarios involving jailbreak and model extraction. Our empirical evaluation shows that Jev remains vulnerable to the examined security and privacy threats, while its decision capability can support beneficial and malicious uses.
Oct 2, 2026

OLMo-Detect: A Multi-Stage, Confounder-Controlled Benchmark for Membership Inference on Large Language Models

Membership inference on large language models (LLMs) aims to determine whether a given text sample was included in an LLM's training data, without access to its training corpus. Despite recent progress, existing benchmarks suffer from three limitations: limited coverage of training stages, insufficient distributional alignment between members and non-members, and lack of rigorous filtering of non-members against the training corpus. To address these limitations, we propose OLMo-Detect, a multi-stage, confounder-controlled benchmark built upon the fully open OLMo 2 pipeline. OLMo-Detect spans pre-training, mid-training, and post-training, explicitly aligns members and non-members on three key axes, and rigorously filters non-members via infini-gram. To assess robustness to distribution shifts, we further introduce OLMo-Detect (Shifted), a variant where members are misaligned with non-members. We evaluate 15 unsupervised and 3 supervised membership inference attacks (MIAs) across the OLMo 2 family, finding that: (i) overall performance is limited: the best unsupervised and supervised MIAs both reach an AUC of only 0.68, and supervised MIAs degrade under cross-domain evaluation; (ii) MIA performance peaks at mid-training and is lower at pre-training and post-training, a pattern driven by data type rather than a stage effect: curated math data is far more detectable than other types; (iii) overall scores improve from 1B to 13B but plateau at 32B; and (iv) no unsupervised MIA is robust to distribution shifts, with AUCs shifting by up to 0.42. Finally, we find that our findings on OLMo 2 generalize to OLMo 3 and non-OLMo models.
Oct 2, 2026

Kinematics-Induced Multimodal 3D Human Pose Estimation with Subject-Level Privacy

Multimodal 3D Human Pose Estimation (3D HPE) combines complementary information from RGB, LiDAR, and mmWave radar, but models trained on correlated observations from the same individuals, raise privacy risks overlooked by record level analysis. We present a unified framework for multimodal 3D HPE that couples kinematics-induced sensor fusion with subject level privacy auditing and private training. First, our multimodal model aligns modality specific joint representation, injects skeletal structure and adaptively aggregates complementary sensor evidence for accurate pose prediction. Second, we formulate a black-box subject membership inference attack for 3D HPE, complemented by an empirical pointwise maximal leakage analysis, which characterizes how individual attack score outcomes change inference about the membership outcome. Third, we instantiate user-level differential privacy via Action Temporal Stratification, a population weighted within-subject sampling strategy that enforces action and temporal coverage. We evaluate our framework on the MM-Fi dataset across three diverse experimental protocols. Source-code will be released upon acceptance.
Sep 28, 2026

Calibrating One-Round Membership Inference with Neighbors

The state-of-the-art Membership Inference (MI) methods calibrate their signal separately for each example using reference models, auxiliary models trained to exclude the target. This paradigm scales poorly to modern large models, however, whose training is too expensive to replicate. This has motivated one-round settings, where only a single trained model is available; but without reference models the per-example calibration that drives the strongest attacks can no longer be estimated, leaving the membership signal weak. We ask whether neighbors of the target point can recover this calibration without training any additional model. Our key observation is that reference models serve only to reveal how an example behaves under models not trained on it, and that querying the target model on nearby samples yields the same information. We propose two complementary ways to obtain such neighbors, and show that querying them against an early training checkpoint further sharpens the signal. We evaluate across three image classification datasets and three training setups, showing that neighbors yield strong membership signals and competitive attack performance at no additional training cost.
Sep 27, 2026

Leaky Students: Membership Inference against On-Policy Distillation

On-policy distillation (OPD) trains a student to match a teacher's next-token distributions on student-generated trajectories. However, privileged information supplied to the teacher for OPD training may contain sensitive data. Whether the student leaks private information about the records supplied to the teacher during distillation remains poorly understood. To the best of our knowledge, we present the first systematic study of membership inference in this setting. We find that fresh student trajectories expose sparse membership signals that fixed reference-answer losses often miss. These signals are mixed with probability changes caused by training on other records. We introduce Leaky, which samples fresh trajectories from the target model and compares its token log-probabilities with the maximum across matched reference models trained without the candidate records. It applies Leaky ReLU to the resulting gaps, preserving positive gaps and downweighting negative gaps as an approximate correction for incidental positive gaps in non-members. Across fifteen targets spanning mathematics, medical question answering, and code generation, Leaky outperforms all evaluated baselines and achieves mean AUROC 0.875, compared with 0.614 for the strongest baseline on each target in the main evaluation. On the same sampled trajectories, the strongest baseline achieves mean AUROC 0.826. These results show that students trained through OPD can expose the membership of records used for teacher supervision, even when fixed reference-answer losses provide little evidence of membership.
Sep 18, 2026

Detecting Pretraining Data in Large Language Models from a Free-Energy Perspective

Detecting pretraining data in large language models is challenging because high likelihood can reflect either training exposure or strong generalization. In the joint space of prediction loss and predictive entropy, a likelihood-only detector uses a horizontal boundary and can mistake predictable non-members for members. Motivated by this, we introduce an inclined boundary that evaluates prediction loss relative to predictive entropy. Our analysis shows that entropy correction can preserve the expected membership signal while reducing its variance, thereby improving standardized member--non-member separation. We further extend the mean--variance analysis to the more general setting with a nonzero mean entropy gap. Interestingly, this entropy-adjusted score admits a Helmholtz free-energy interpretation, leading to Energy Transfer Detection (ETD), which views pretraining data detection from a macroscopic residual free-energy transfer perspective. Extensive experiments show that ETD achieves the best average detection performance, improving average AUROC by up to 3.5% and TPR@5%FPR by up to 5.1%, while remaining robust across diverse settings.
Sep 11, 2026

Membership Inference via Pairwise Likelihood Ratios

Membership inference attacks (MIAs) are the standard tool for auditing the privacy risks of machine learning models. Given a query point, an MIA aims to determine whether that point was used to train the target model. In practice, such inference must rely on the statistical signals exposed by the model's outputs, such as confidence scores, logits, and intermediate feature representations. However, existing methods often fail to efficiently summarize and combine these statistical signals. To address this limitation, we propose Pairwise Likelihood MIA (PL-MIA), a unified method that combines a Gaussian likelihood-ratio (GLR) statistic with population calibration and the Cauchy combination test. We characterize theoretically how the GLR retains variance-contraction signals and establish conditions under which population calibration and Cauchy combination improve attack power. We obtain pp-values from pairwise comparisons between the query point and reference points not used for training, and aggregate these continuous signals using the Cauchy combination test. This preserves the evidence strength that is discarded when each pairwise comparison is reduced to a binary vote. Extensive experiments demonstrate that PL-MIA outperforms strong baselines, improving the true positive rate (TPR) by over 25% in the critical low-false-positive regime, corroborating our theoretical findings. These results demonstrate how statistical principles can turn noisy model outputs into more powerful, calibrated, and reproducible evidence for membership privacy auditing.
Sep 10, 2026

Predicting Privacy Leakage from Weight Spectral Density

Membership inference attacks (MIAs) are widely used to audit the privacy disclosure risk of machine learning models, however current state-of-the-art attacks require training computationally expensive shadow models, making large-scale privacy evaluation impractical. In this work, we investigate whether inexpensive spectral metrics derived from the heavy-tailed self-regularisation framework can serve as proxies for MIA vulnerability. We evaluate several WeightWatcher spectral metrics on image and tabular classification tasks and compare their relationship with MIA privacy leakage against conventional measures of generalisation. Across datasets, stable rank exhibits a strong positive correlation with overall MIA success, while Log alpha-Norm shows a consistent negative correlation with MIA vulnerability at the low false-positive regime. These associations are observed to be stronger than those obtained using the generalisation gap. The results indicate that neural network spectra may contain information about privacy leakage that is not fully captured by conventional measures of overfitting, motivating spectral analysis as a promising direction for scalable privacy auditing.
Sep 10, 2026

Empirical Evaluation of Membership Inference Attacks on NLP Text Classifiers: A Baseline Study on SST-2

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.