The paper investigates whether large, instruction‑following Vision‑Language Models (VLMs) can reliably perform zero‑shot image privacy classification. It compares three open‑source VLMs to specialized privacy models on two public benchmarks, evaluating accuracy, robustness to image degradations (compression, lighting changes, noise), inference speed, and parameter count. The findings show that while VLMs remain robust to perturbations, they are less accurate and significantly slower than smaller, purpose‑built privacy models, indicating that scaling alone does not guarantee effective privacy classification.
By Alina Elena Baia, Alessio Xompero, Andrea Cavallaro
The paper introduces XAVAD, a unified approach for explainable anomaly detection in anonymized crowd video monitoring. It combines zero‑shot anonymization, anomaly type categorization, spatial localization, and scene context description using multimodal CLIP embeddings, LLM‑assisted prompting, and a prompt‑guided patching autofocus mechanism. Experiments on public CCTV datasets show XAVAD outperforms state‑of‑the‑art methods and offers insights into privacy‑explainability trade‑offs.
By Mulugeta Weldezgina Asres, Lei Jiao, Christian Walter Omlin
The paper compares four audit methods for assessing identity‑level differential privacy in pre‑trained, black‑box face generators. Each method—GaussMech, KDE‑LR, MMD‑TV, and ROC‑HT—has distinct assumptions, hyperparameters, and finite‑sample limitations, and they produce markedly different epsilon estimates when applied to FaceFusion and InstantID. The study finds that all methods reveal significant identity distinguishability, but none can be reliably ranked in this high‑distinguishability regime, suggesting that future work should evaluate them on partially private mechanisms.
By Arman Zareian Jahromi, Vishnu Bondalakunta, Mohammad Akbar Bin Shah, Naimul Haque, Shuangqing Wei, George T. Amariucai
arXiv:2608.28691v1 Announce Type: cross
Abstract: Wearable VLM pipelines promise continuous multimodal assistance from egocentric visual capture: a user asks a task-driven question about the surround...
By Zhimin Li, Pan Wang, Jingxian Chen, Yuantao Tang, Anthony Chen, Qian Lou, Jingtong Hu
The paper argues that evaluating privacy‑enhancing technologies (PETs) solely through image classification is insufficient because classification remains robust to many geometric and local perturbations. It proposes a compute‑aware multi‑task protocol that uses lightweight proxy tasks to assess PETs across various transformations, revealing that PETs with similar classification accuracy can perform very differently on other vision tasks. The study demonstrates the necessity of broader evaluation metrics beyond classification to truly gauge PET effectiveness.
By Leon Ranke, Wolfgang H\"ubner, Ronny Hug, Michael Arens, J\"urgen Beyerer
arXiv:2606. 09132v1 Announce Type: new Abstract: Visual Language Models (VLMs) have gained significant popularity due to their remarkable ability.
By Tiejin Chen, Pingzhi Li, Kaixiong Zhou, Tianlong Chen, Hua Wei
The paper introduces a unified evaluation framework for face de-identification (FDeID), addressing inconsistencies in current protocols. It presents UtilFace, a demographically balanced benchmark assembled from four large-scale datasets, and HiFD, a hierarchical metric that simultaneously measures identity suppression, multi-level utility preservation, and image quality. Using this framework, the authors compare adversarial, GAN-based, and diffusion-based methods, revealing trade-offs and failure modes that were previously hidden, and they release the benchmark and toolkit for reproducible research.
By Hui Wei, Hao Yu, Hui Kuurila-Zhang, Guoying Zhao
arXiv:2606. 06784v1 Announce Type: cross Abstract: Public social media posts can reveal private information through weak cues scattered across text, images, or metadata.
By Zifan Peng, Yini Huang, Aiwen Lu, Qiming Ye, Peixian Zhang, Jingyi Zheng, Yule Liu, Xuechao Wang, Xinlei He, Jiaheng Wei
The paper introduces XCal-FL, a federated learning algorithm that dynamically calibrates differential privacy noise using three signals—prediction logit variations, counterfactual margins, and saliency concentration—to improve both predictive accuracy and explanation fidelity. Experiments on medical imaging datasets demonstrate that XCal-FL outperforms static-noise and state‑of‑the‑art adaptive DP methods, achieving over 10% better accuracy and up to fivefold higher explanation fidelity while using privacy budgets more efficiently. The study highlights that explanation fidelity behaves non‑linearly with privacy loss, indicating that explainability is a separate dimension of the privacy trade‑off.
By Michael Khavkin, Kichang Lee, Jaeho Jin, JeongGil Ko, Eran Toch
Hide and Seek (HnS) is a concept‑level attribution method that interprets black‑box vision models without needing model weights or labeled concept datasets. It retrieves concept‑aligned images via a vision‑language model, localizes and masks the concept region, and measures attribution by observing changes in the model’s output. HnS was validated on CelebA and COCO‑Stuff, showing stronger agreement with ground‑truth attributions than existing methods, and was applied to tasks like image captioning and satellite image segmentation where other concept attribution methods fail.
By Bhanu Tokas, Rahul Nair, Rockwell Jackson, Hannah Kerner
The paper introduces XCal-FL, a federated learning algorithm that dynamically calibrates differential privacy noise using three signals—prediction logit variations, counterfactual margins, and saliency concentration—to improve both predictive accuracy and explanation fidelity. Experiments on medical imaging datasets demonstrate that XCal-FL outperforms static-noise and state‑of‑the‑art adaptive DP methods, achieving over 10% better accuracy and up to five‑fold higher explanation fidelity while using privacy budgets more efficiently. The study reveals that explanation fidelity behaves non‑linearly with privacy loss, indicating that explainability is a separate dimension of the privacy trade‑off that cannot be inferred from utility alone.
arXiv:2510.24233v2 Announce Type: replace
Abstract: Deep generative models are often trained on sensitive data, such as genetic sequences, health data, or more broadly, any copyrighted, licensed or p...
By Antoine Szatkownik (TAU, BioInfo), Aur\'elien Decelle (TAU), Beatriz Seoane (TAU), Nicolas B\'ereux (TAU), L\'eo Planche (BioInfo), Guillaume Charpiat (TAU), Burak Yelmen (BioInfo, TAU), Flora Jay (BioInfo, TAU), Cyril Furtlehner (TAU)