Research map / Adversarial robustness and security / Backdoor and poisoning
Indelible Backdoors: On the Limits of Post-Training Defenses
Jingyi Guo, Thuy Dung Nguyen, Taylor T Johnson, Kevin Leach
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 |
|---|---|---|---|
| align backdoor and clean gradients on Fisher-important neurons | trigger generator backdoor attack · CIFAR-10 under FST | ASR: COMBAT 30.65 -> Ours 99.58 | Improved |
| use gradient-aligned backdoor against fine-tuning defenses | backdoor attack · CIFAR-10 under FT/NAD/FST/TSBD | ASR: ASR above 99.5% under all four defenses | Improved |
| pruning defense ANP applied to the aligned backdoor | Ours · CIFAR-10 under ANP | ASR: 100.00 -> 46.50 | Hurt |
| increase perturbation budget epsilon from 0.03 to 0.1 (ablation) | Ours · CIFAR-10 under TSBD | ASR: 47.76 -> 99.84 | Improved |
| change perturbation budget on clean accuracy (ablation) | Ours · CIFAR-10 pretrained | clean accuracy: 92.82 to 93.27 across epsilon 0.075-0.1; 92.91 at 0.03 | No clear effect |
| lower poisoning ratio from 0.1 to 0.01 (ablation) | Ours · CIFAR-10 under defenses | ASR: FST 99.58 -> 99.29 | No clear effect |
| lower poisoning ratio from 0.1 to 0.01 (ablation) | Narcissus · CIFAR-10 pretrained | ASR: 94.64 -> 38.26 | Hurt |
| reduce defender fine-tuning data ratio (ablation) | defenses vs Ours · CIFAR-10 | ASR: I-BAU 99.63 -> 31.72 at ratio 0.01; FST 99.46 -> 100.00 | Mixed |
| evaluate on ViT-B/16 instead of PreAct-ResNet18 (ablation) | Ours · CIFAR-10 under FT, TSBD | ASR: FT 99.99 -> 10.28; TSBD 99.84 -> 14.04 | Hurt |
| evaluate on Tiny-ImageNet vs CIFAR-10 | Ours under FST · Tiny-ImageNet | ASR: 99.58 -> 21.38 with ACC 32.74 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- COMBAT — COMBAT: Alternated Training for Effective Clean-Label Backdoor Attacks (AAAI 2024) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Enhancing Fine-Tuning Based Backdoor Defense with Sharpness-Aware Minimization (ICCV 2023 · similarity 0.93)
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.91)
- A Theoretical Analysis of Backdoor Learning as Simplicity-Biased Optimization Dynamics (NeurIPS 2026 · similarity 0.90)
- Uncovering, Explaining, and Mitigating the Superficial Safety of Backdoor Defense (NeurIPS 2024 · similarity 0.88)
- Backdoor Mitigation via Invertible Pruning Masks (NeurIPS 2025 · similarity 0.88)
- Fusing Pruned and Backdoored Models: Optimal Transport-based Data-free Backdoor Mitigation (AAAI 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.
