Research map / Continual learning / Continual fine-tuning of LLMs
Static and Sequential Malicious Attacks in the Context of Selective Forgetting
CHENXU ZHAO, Wei Qian, Zhitao Ying, Mengdi Huai
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 |
|---|---|---|---|
| use the proposed optimization instead of random removal (RandSearch) | static forgetting attack, first-order · Diabetes targeted | attack success rate: 0.04 ± 0.03 -> 1.00 ± 0.00 | Improved |
| use the proposed optimization instead of random removal (RandSearch) | static forgetting attack, first-order · CIFAR-10 targeted | attack success rate: 0.08 ± 0.04 -> 0.80 ± 0.04 | Improved |
| use the proposed optimization instead of RandSearch against SISA | static forgetting attack, SISA · Diabetes targeted | attack success rate: 0.47 ± 0.12 -> 0.74 ± 0.07 | Improved |
| use the proposed optimization instead of RandSearch against SISA | static forgetting attack, SISA · CIFAR-10 targeted | attack success rate: 0.13 ± 0.05 -> 0.67 ± 0.08 | Improved |
| apply optimization framework when unlearned data is underrepresented (1% to 5% of CIFAR-10) (ablation) | first-order and second-order unlearning · CIFAR-10 | attack success rate | Improved |
| increase the percentage of unlearning samples from the target class (5% to 25%) | static forgetting attack · Diabetes | attack success rate | Improved |
| attack a subpopulation of size 50 with unrolling SGD | static forgetting attack · Diabetes targeted | attack success rate: 1.0 | No clear effect |
| attack a subpopulation of size 50 with amnesiac unlearning | static forgetting attack · CIFAR-10 targeted | attack success rate: 0.85 | Hurt |
| choose subpopulation members by feature similarity instead of randomly (ablation) | static forgetting attack · Diabetes and CIFAR-10 subpopulations | attack success rate | Improved |
| unlearn points from less redundant clusters (10% sampled) instead of more redundant (50%) (ablation) | static forgetting attack · Diabetes and CIFAR-10 | attack success rate | Improved |
| use optimization over random selection of points within each cluster (ablation) | static forgetting attack · Diabetes and CIFAR-10 diversity study | attack success rate | Improved |
| apply the learned adversarial policy instead of random or no attack | sequential forgetting attack, first-order · Synthetic, Adult, MNIST | Euclidean distance to target model | Improved |
Closest papers by meaning
- Backdoor Attacks via Machine Unlearning (AAAI 2024 · similarity 0.88)
- System-Aware Unlearning Algorithms: Use Lesser, Forget Faster (ICML 2025 · similarity 0.87)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.87)
- Rethinking Adversarial Robustness in the Context of the Right to be Forgotten (ICML 2024 · similarity 0.87)
- Fast Machine Unlearning without Retraining through Selective Synaptic Dampening (AAAI 2024 · similarity 0.87)
- Machine Unlearning Fails to Remove Data Poisoning Attacks (ICLR 2025 · similarity 0.87)
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.
