Research map / Federated learning and privacy / Privacy attacks and secure learning
Reminiscence Attack on Residuals: Exploiting Approximate Machine Unlearning for Privacy
Yaxin Xiao, Qingqing Ye, Li Hu, Huadi Zheng, Haibo Hu, Zi Liang, Haoyang Li, Yijie Jiao
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 the replay phase (two-phase design) to orthogonal unlearning (ablation) | OUR · CIFAR20 ablation | ReA attack accuracy: ReA accuracy 95.52% without replay; unlearning efficacy degraded by 0.06 | Improved |
| use orthogonal unlearning instead of random labels in the first phase (ablation) | OUR (with Replay) · CIFAR classification | ReA attack accuracy: 2.89-6.12% lower ReA accuracy | Improved |
| add l1 regularization in the replay phase (ablation) | OUR · class-wise unlearning | ReA attack accuracy: 35.07% lower ReA accuracy | Improved |
| add l1 regularization in the replay phase (ablation) | OUR · sample-wise unlearning | ReA attack accuracy | No clear effect |
| select first, middle and last block outputs (Span 3) for orthogonal layers (ablation) | OUR · CIFAR20 | ToW / ReA accuracy | Improved |
| aggregate multiple learning rates in ReA instead of a single lr (ablation) | ReA · CIFAR20 class-wise MIA | ReA attack accuracy | Improved |
| raise ReA learning rate to 0.1 (ablation) | ReA · CIFAR20 class-wise MIA | ReA attack accuracy | Hurt |
| set ReA convergence threshold between 70% and 80% (ablation) | ReA · CIFAR20 class-wise MIA | ReA attack accuracy: 70% and 80% | Improved |
| add SalUn parameter freezing to OUR (SalUn-OUR) (ablation) | OUR · identity unlearning (Stable Diffusion) | RISM: 0.68 -> 0.74 | Improved |
| use ReA instead of MIA-LiRA against random-label unlearning | membership inference · CIFAR20 (-5) class-wise | balanced attack accuracy: 50.00 -> 95.52 | Improved |
| apply OUR instead of RUM | approximate unlearning · CIFAR20 (-5) class-wise | ReA balanced accuracy: 85.64 -> 52.28 | Improved |
| use DiffAtk instead of ReA against SalUn-unlearned SD | attack · identity unlearning | resonance index: ReA 2 vs DiffAtk 10 (ISM 0.77 vs 0.10) | Hurt |
Closest papers by meaning
- Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable (NeurIPS 2024 · similarity 0.91)
- Leveraging Per-Instance Privacy for Machine Unlearning (ICML 2025 · similarity 0.91)
- The Utility and Complexity of In- and Out-of-Distribution Machine Unlearning (ICLR 2025 · similarity 0.90)
- Certified Unlearning for Neural Networks (ICML 2025 · similarity 0.89)
- The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed Samples (NeurIPS 2025 · similarity 0.89)
- Unlearned but Not Forgotten: Data Extraction after Exact Unlearning in LLM (NeurIPS 2025 · similarity 0.89)
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.
