Research map / Adversarial robustness and security / Certified and empirical robustness
A Unified Perspective on Adversarial Membership Manipulation in Vision Models
Ruize Gao, Kaiwen Zhou, Yongqiang Chen, Feng Liu
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 momentum to cosine-annealed step size in MFA (ablation) | MFA · CIFAR-10 (eps 4/255) | Error Area: 0.9444 -> 0.9451 | Improved |
| replace fixed step size with cosine annealing plus momentum in MFA (ablation) | MFA · CIFAR-100 (eps 4/255) | Error Area: 0.5917 -> 0.6136 | Improved |
| replace halving heuristic (APGD) with cosine annealing plus momentum in MFA (ablation) | MFA · CINIC-10 (eps 4/255) | EER: 85.00% -> 86.85% | Improved |
| add momentum on top of cosine annealing (ablation) | MFA · CINIC-10 (eps 4/255) | EER: 86.75% -> 86.85% | No clear effect |
| increase perturbation budget epsilon for gradient-norm detection (ablation) | MFD · CINIC-10 | detection AUC: 0.6684 -> 0.9935 (eps 1/255 -> 8/255) | Improved |
| add gradient-norm weighting to Attack R | Attack R · SVHN (with fabricated members) | AUC: 0.3172 -> 0.7185 (lambda 5) | Improved |
| add gradient-norm weighting to LiRA | LiRA · CINIC-10 (with fabricated members) | AUC: 0.6734 -> 0.8166 (lambda 5) | Improved |
| add gradient-norm weighting to RMIA | RMIA · CIFAR-10 (with fabricated members) | AUC: 0.6834 -> 0.8065 (lambda 35) | Improved |
| increase lambda in gradient-norm weight (ablation) | AR-Attack R / AR-LiRA · SVHN / CINIC-10 | AUC: 0.7185 -> 0.5660 (SVHN); 0.8166 -> 0.7872 (CINIC-10) | Hurt |
| increase lambda in gradient-norm weight (ablation) | AR-RMIA · CIFAR-10 | AUC: 0.7834 -> 0.8065 | Improved |
Closest papers by meaning
- Membership Inference Attacks against Large Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- Black-Box Membership Inference Attack for LVLMs via Prior Knowledge-Calibrated Memory Probing (NeurIPS 2025 · similarity 0.87)
- Assessing Per-Sample Membership Inference Vulnerability without Retraining (NeurIPS 2026 · similarity 0.86)
- OSLO: One-Shot Label-Only Membership Inference Attacks (NeurIPS 2024 · similarity 0.86)
- Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models (CVPR 2025 · similarity 0.86)
- Theoretically Unmasking Inference Attacks Against LDP-Protected Clients in Federated Vision Models (ICML 2025 · similarity 0.85)
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.
