Research map / Federated learning and privacy / Privacy attacks and secure learning
Compositional Non-Face Re-Identification Pressure under Cumulative Vision Releases
Tirth Joshi, Honggang Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add release R2 (encoder trained on full exposure) | A0 retrieval attacker · Market-1501-Tau (N=751) | Pguess: 0.2883 -> 0.7491 | Improved |
| add release R2 (encoder trained on full exposure) | A0 retrieval attacker · Market-1501-Tau (N=751) | vRPI2: 4.3319 -> 6.2119 | Improved |
| add release R3 (generic ImageNet ResNet-50 encoder) with calibrated fusion | A0 retrieval attacker · Market-1501-Tau (N=751) | Pguess: 0.7491 -> 0.7529 | No clear effect |
| add release R3 (generic ImageNet ResNet-50 encoder) with calibrated fusion | A3 MLP probe · Market-1501-Tau (N=751) | Pguess: 0.7063 -> 0.7092 | No clear effect |
| add R4 prototype bank released, exploited by prototype-index attacker | A2 · Market-1501-Tau (N=751) | Pguess: Pguess=0.7519 (vRPI2=6.3572) | Improved |
| add R4 kNN graph released, exploited by graph attacker | A0-Graph · Market-1501-Tau (N=751) | Pguess: 0.7366 (vRPI2=5.6697) | Improved |
| add R2 exposure expansion on CUHK03-Tau | A0 · CUHK03-Tau (paired protocol, N=700) | vRPI2: .530 -> 3.325 | Improved |
| add R2 exposure expansion on CUHK03-Tau | A3 · CUHK03-Tau (paired protocol, N=700) | vRPI2: 4.549 -> 4.977 | Improved |
| add R2 exposure expansion on Market-1501-Tau paired protocol | A0 · Market-1501-Tau (paired protocol, N=736) | Pguess: .197 -> .547 | Improved |
| increase top-region masking fraction from 0 to 40% (ablation) | A0 · Market-1501-Tau masking sweep | vRPI2: stays in [4.87, 4.91] | No clear effect |
| increase top-region masking fraction from 0 to 40% (ablation) | A3 · Market-1501-Tau masking sweep | vRPI2: stays in [5.83, 5.88] | No clear effect |
| sweep posterior temperature from .025 to 1.0 (ablation) | A0 · Market-1501-Tau | vRPI2: A0 spanning [.781, 4.100] | Mixed |
Closest papers by meaning
- From Weak Cues to Real Identities: Evaluating Inference-Driven De-Anonymization in LLM Agents (ICML 2026 · similarity 0.87)
- Disguise without Disruption: Utility-Preserving Face De-identification (AAAI 2024 · similarity 0.86)
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.86)
- Unifying Re-Identification, Attribute Inference, and Data Reconstruction Risks in Differential Privacy (NeurIPS 2025 · similarity 0.86)
- PerceptAnon: Exploring the Human Perception of Image Anonymization Beyond Pseudonymization for GDPR (ICML 2024 · similarity 0.85)
- Auditing Privacy Mechanisms via Label Inference Attacks (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
