Research map / Adversarial robustness and security / Backdoor and poisoning
Adversarial Unlearning of Backdoors via Implicit Hypergradient
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 I-BAU instead of the strongest baseline, TABOR | simplified VGG · CIFAR-10, BadNets one-trigger | ASR: 8.12 -> 12.30 | Mixed |
| Use I-BAU instead of Deepinspect | simplified VGG · GTSRB, BadNets one-trigger | ASR: 0.47 -> 3.32 | Hurt |
| Apply I-BAU backdoor unlearning | simplified VGG · GTSRB, 7-trigger-7-target | avg. ASR: 99.37 -> 4.18 | Helped |
| Apply I-BAU with only 100 clean samples | simplified VGG · CIFAR-10, Trojan WM | ASR: 99.96 -> 4.00 | Helped |
| Reduce available clean data from 2,500 to 100 samples | I-BAU · CIFAR-10, Trojan WM | ASR: 6.96 -> 4.00 | Hurt |
| Reduce poison ratio from 5.0% to 0.5% | I-BAU · CIFAR-10, Trojan WM | ASR: 9.78 -> 13.08 | Mixed |
| Use I-BAU rather than Fine-pruning | simplified VGG · CIFAR-10, one-trigger-one-target | runtime (s): 45.33 -> 6.82 | Helped |
| Use I-BAU rather than NAD | simplified VGG · GTSRB, one-trigger-one-target | runtime (s): 79.14 -> 7.84 | Helped |
| Increase the I-BAU norm bound (ablation) | I-BAU · clean accuracy (Appendix experiment) | ACC | No clear effect |
| Optimize a shared universal trigger and account for its dependence on model parameters using an implicit hypergradient | backdoor unlearning · CIFAR-10 and GTSRB backdoor evaluations | ASR | Helped |
Closest papers by meaning
- Shared Adversarial Unlearning: Backdoor Mitigation by Unlearning Shared Adversarial Examples (NeurIPS 2023 · similarity 0.92)
- Anti-Backdoor Learning: Training Clean Models on Poisoned Data (NeurIPS 2021 · similarity 0.90)
- Adversarial-Inspired Backdoor Defense via Bridging Backdoor and Adversarial Attacks (AAAI 2025 · similarity 0.90)
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.89)
- RPP: A Certified Poisoned-Sample Detection Framework for Backdoor Attacks under Dataset Imbalance (NeurIPS 2026 · similarity 0.89)
- BackdoorDM: A Comprehensive Benchmark for Backdoor Learning on Diffusion Model (NeurIPS 2025 · similarity 0.88)
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.