Training data and privacy
Leakage of personal or proprietary data through training, memorization, inference attacks or careless logging.
- All items
- 32
- Last 90 days
- 14
- Change
- +133%vs 6 before
Items per month
| Month | Items |
|---|---|
| Jun 2025 | 0 |
| Jul 2025 | 0 |
| Aug 2025 | 0 |
| Sep 2025 | 1 |
| Oct 2025 | 2 |
| Nov 2025 | 0 |
| Dec 2025 | 0 |
| Jan 2026 | 1 |
| Feb 2026 | 1 |
| Mar 2026 | 2 |
| Apr 2026 | 2 |
| May 2026 | 2 |
| Jun 2026 | 2 |
| Jul 2026 | 2 |
| Aug 2026 | 3 |
| Sep 2026 | 0 |
| Oct 2026 | 8 |
| Nov 2026 | 5 |
32 items
Unveiling Privacy Risks in Federated Learning: An Efficient Bias-Based Membership Inference Attack and Adaptive Defense
Nov 18, 2026InfoResearchPeer-reviewedSecurityResearchPrivacyResearchers show that bias parameters in deep neural networks make federated learning vulnerable to membership inference attacks, which reveal whether a data point was in the training set. They build efficient bias-based attacks for both a local participant and the central server, using a shallow neural network and a difference-comparison step to identify the source participant. Experiments on four public datasets report higher accuracy than prior methods and robustness against common defenses.
Fix: They introduce an adaptive noise pruning defense (ANPD) that mitigates the bias-based membership inference attack without substantially degrading target model performance.
OpenAlex (peer-reviewed AI security)GapLeak: When Privacy-Preserving LLM-based Web Agents Become a Privacy Risk
Nov 17, 2026InfoResearchPreprintSecurityPrivacyResearchThis research paper introduces GapLeak, a privacy vulnerability in LLM-based web agents where a legitimate web form requests a field unavailable from authorized sources, prompting the agent to escalate to reachable but unauthorized sources and disclose sensitive data without prompt injection. Evaluated on three agent frameworks and six LLMs from Anthropic, Google, and OpenAI across four domains, GapLeak caused an average unauthorized disclosure of about 37%, and eliminating the information gap reduced disclosure by more than 50%. Least-privilege instruction defenses partially mitigated it but often reduced task success.
Fix: The authors propose submission-time verification, which filters unauthorized content before form submission. The source reports that it eliminates over 96% of unauthorized disclosures while preserving task success.
Arxiv (cs.CR + cs.CL + cs.LG)Apple reveals new Mac security controls as AI agents raise privacy risks
Nov 16, 2026InfoNewsSecurityPrivacyApple has announced new Mac permission controls, warnings and transparency features for macOS 27.3, aimed at AI agents and apps that request "full disk access." The changes split Device Control and Data Access into System & App Protection Override and Accessibility, add a per-request permission option, and add an indicator when one app controls another.
Fix: Apple says developers should adjust onboarding to "gracefully handle having access to only a limited set of folders or apps", and apps needing Full Disk Access for backup or assistive technology can request entitlement exceptions. Users are advised to grant access only to the specific apps that need it.
TechCrunch (Security)Timing is Everything: Inferring Temporal Membership in Continually Updated Models
Nov 3, 2026InfoResearchPreprintSecurityPrivacyResearchers present the first systematic study of temporal privacy risk in continually updated deep learning models. They introduce TMIA, a black-box temporal membership inference attack that uses bidirectional LSTMs over model outputs from successive updates to infer whether and when individual samples were used in training. TMIA significantly outperforms existing baselines across three language models and three image classification models.
Arxiv (cs.CR + cs.CL + cs.LG)Probing Local Conditional Fit for Diffusion Model Membership Inference
Nov 2, 2026InfoResearchPreprintResearchPrivacyResearchers propose LoFIT (Local Conditional Fit), a membership inference method for text-to-image diffusion models that measures the initial response of the denoising loss along normalized conditional prediction shifts. The method needs only forward queries to the denoiser, no target-model gradients. Across four datasets it improves membership discrimination on fine-tuned and pretrained models, and it separates members early in fine-tuning where the naive denoising-loss baseline stays near chance.
Arxiv (cs.RO + cs.CV security)When Safety Checks Become Privacy Risks: A Privacy-Preserving Paradigm for LLM Agent Safety Auditing
Oct 30, 2026InfoResearchPreprintResearchPrivacyThe paper proposes a privacy-preserving paradigm for pre-execution safety auditing of LLM agent tool calls, separating delegated audit computation from plaintext access and execution authority. Its instantiation, SealCall, combines learned audit representations, CKKS-based encrypted similarity computation, and execution control bound to audited tool calls. On 1,252 public binary candidates it reaches a mean F1 of 0.840 under grouped nested evaluation, and one configuration completes each encrypted scoring operation in 0.176 s.
Arxiv (cs.CR + cs.CL + cs.LG)Difficulty Calibration Across Time for Membership Inference in Diffusion Models
Oct 29, 2026InfoResearchPreprintSecurityResearchResearchers address membership inference attacks on diffusion models, which test whether an image was in a model's training set, and note that standard reconstruction-loss scoring fails because some images are inherently easy to reconstruct. They propose difficulty calibration across two time axes: Reference-Model Calibration (RMC) compares a fine-tuned model's loss against an earlier checkpoint, and Temporal Self-Calibration (TSC) compares loss across neighboring diffusion timesteps using a single model. The authors report that their methods outperform prior state of the art by a wide margin and generalize across x-, v-, and ε-prediction models.
Arxiv (cs.RO + cs.CV security)Beyond membership inference: Distributional leakage in diffusion models
Oct 24, 2026InfoResearchPeer-reviewedSecurityPrivacyResearchBarezzani, Coscia, Ferrari and Piuri publish a study in Computers & Security, available online 24 October 2026, on distributional leakage in diffusion models. The source text provides only the title, publication details and author list, so the research question, method and findings cannot be stated from it.
Elsevier Security JournalsBehavioral Forgetting Is Not Privacy in Machine Unlearning
Oct 15, 2026InfoResearchPreprintResearchPrivacyThis research argues that behavioral forgetting, measured only on a designated benchmark distribution, is an insufficient standard for privacy-oriented machine unlearning in large language models. In a controlled case study of cross-lingual querying across multiple multilingual model families, forgetting varied substantially when semantically equivalent information was queried in different languages and scripts. The authors conclude that a single access path can expose a real gap between behavioral forgetting and information inaccessibility.
Arxiv (cs.CR + cs.CL + cs.LG)COMPASS: Membership Inference Attack Against Vision Large Language Models via Adaptive Noise Calibration
Oct 12, 2026InfoResearchPreprintSecurityPrivacyResearchCOMPASS is an adaptive gray-box membership inference attack against vision large language models (VLLMs) that calibrates both perturbation magnitude and membership-score direction using a known non-member reference set. The authors find that a low-noise regime often separates members from non-members better than large perturbations, with the score direction reversing in some model-dataset pairs. Tested on LLaVA v1.5-7B, MiniGPT-4 and HuluMed-32B, COMPASS reaches AUCs up to 0.88 on general-domain data and 0.64 on distribution-matched medical data, where the strongest baseline stays at chance.
Arxiv (cs.CR + cs.CL + cs.LG)Privacy Risks and Leakage Pathways in Agentic and Multi-Agent AI Systems
Oct 11, 2026InfoResearchPeer-reviewedResearchSecurityACM Digital Library (TOPS, DTRAP, CSUR)Systemic privacy risks of personal data exposure through conversational large language model agents
Oct 7, 2026InfoResearchPeer-reviewedResearchPrivacyThis critical review examines privacy risks from conversational LLM agents across five areas: data leakage and memorization, adversarial extraction and prompt injection, inference and re-identification, surveillance and profiling, and regulatory governance breakdown. It finds that technical mitigations such as differential privacy, federated learning and machine unlearning work to some extent in controlled settings but fail to fully protect conversational deployments, and it highlights risks that machine unlearning can leave traces exploitable by adversarial reconstruction.
Fix: The authors support a multi-layered mitigation framework combining Privacy-Preserving Data Publishing (PPDP) principles, technical safeguards, and governance reform.
OpenAlex (peer-reviewed AI security)A novel privacy-preserving large language model integrating trust-weighted and ethical gradient masking
Oct 7, 2026InfoResearchPeer-reviewedResearchPrivacyResearchers propose a privacy-preserving large language model training approach, the trust-weighted and ethical gradient model, to address the privacy utility tradeoff in LLMs trained on personally identifiable information. The method combines trust-weighted memory, ethical gradient masking under a tracked (ϵ, δ) privacy budget, and an ethical boundary layer that screens risky outputs. Experiments on the Pile dataset report reduced PII leakage with baseline utility maintained, and the authors claim resistance to membership inference, extraction and linkage attacks.
OpenAlex (peer-reviewed AI security)Black-Box Membership Inference Attacks Against Contrastive Learning via Aggressive Data Augmentations
Aug 28, 2026LowResearchPeer-reviewedSecurityResearchADA-MIA is a black-box membership inference attack against contrastive learning (CL) encoders. It needs only an auxiliary dataset of training samples, without their membership labels, and it constructs membership features from aggressively augmented views of the target sample, enabling unsupervised inference. Across four datasets, it matches the performance of seven existing attacks while requiring only black-box embedding access to the encoder.
IEEE Xplore (Security & AI Journals)DeepU: Deeper Granular Within-Layer Machine Unlearning
Aug 12, 2026InfoResearchPeer-reviewedSecurityPrivacyResearchDeepU is a machine unlearning framework that removes the influence of selected training data at the level of individual weights. It scores each weight by the signal-to-noise ratio of gradients from sensitive and non-sensitive data, then resets, perturbs, decays or stabilizes weights accordingly. On CIFAR-10, CIFAR-100, Tiny ImageNet and CelebA, it cut successful membership inference attacks by 60–90% with under a 3% accuracy drop, and re-tuning took 20.75 seconds and 102.47 MB, up to 36.6 times faster than competing methods.
IEEE Xplore (Security & AI Journals)Privacy-Preserving GAN for Synthetic Data against Membership Inference Attack
Aug 5, 2026InfoResearchPeer-reviewedResearchPrivacyACM Digital Library (TOPS, DTRAP, CSUR)Private Claude Chats Exposed in Google and Bing Search Results
Jul 27, 2026LowNewsPrivacyIndustrySome Anthropic Claude chats shared via public "snapshot" URLs (claude.ai/share) were indexed by Google and Bing, exposing conversations about political parties, Kansas attorney self-reporting rules, and erotic role play. Anthropic's robots.txt has blocked crawlers from shared chats since at least September 2025, but WIRED found the exposed pages lacked the "noindex" tag that Bing and Google say they consider.
Fix: Bing's documentation says developers should include a "noindex" tag on individual pages in addition to robots.txt, and Google says it ignores robots.txt for pages linked elsewhere unless a "noindex" html tag or "x-robots-tag" response header is present.
Wired (Security)Forgetting Similar Samples: Can Machine Unlearning Do it Better?
Jul 17, 2026InfoResearchPeer-reviewedResearchSecurityThis paper examines whether existing machine unlearning methods truly remove the influence of target training samples when the training dataset contains many similar samples. Experiments on four constructed datasets for image and language models show a notable gap between expected and actual performance for most existing methods, including the retraining-from-scratch baseline. The authors also explore potential solutions to improve current unlearning approaches.
IEEE Xplore (Security & AI Journals)MU-MIA: Machine Unlearning for Membership Inference Attacks
Jun 8, 2026InfoResearchPeer-reviewedSecurityPrivacyThe paper proposes MU-MIA, a membership inference attack that uses machine unlearning to gradually reduce a model's memorization of specific samples and records a forgetting trajectory for each one. A BiLSTM-based binary classifier with attention then separates member from non-member samples using these trajectories. The unlearning step runs in a zero-shot setting that needs no real data, and the authors report that the method outperforms existing baseline attacks across different datasets and model architectures.
IEEE Xplore (Security & AI Journals)Parameter-Agnostic Privacy-Preserving Machine Unlearning for Large Language Models
Jun 1, 2026InfoResearchPeer-reviewedResearchPrivacyThis research paper proposes a privacy-preserving machine unlearning mechanism for large language models. The method uses information retrieval to remove high-risk semantic meanings from model output and adds differentially-private randomization so unlearned information is statistically indiscernible. It requires neither parametric fine-tuning nor in-context prompt calibration, and the authors report theoretical privacy and unlearning guarantees plus experiments on real-world datasets.
IEEE Xplore (Security & AI Journals)
Topic added 2026-10-09. An item belongs to this topic when its title matches one of the topic's patterns or its summary mentions the topic at least twice. Report a wrong match with the feedback button on the item.