Research map / Adversarial robustness and security / Backdoor and poisoning
Backdoor Mitigation by Distance-Driven Detoxification
Shaokui Wei, Jiayin Liu, Hongyuan Zha
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 D3 instead of vanilla fine-tuning | PreAct-ResNet18 backdoored model · CIFAR-10 avg of 7 attacks | ASR: 38.86 -> 0.46 | Improved |
| use D3 instead of FT-SAM | PreAct-ResNet18 backdoored model · CIFAR-10 avg of 7 attacks | ASR: 3.91 -> 0.46 | Improved |
| use D3 instead of SAU | PreAct-ResNet18 backdoored model · CIFAR-10 avg of 7 attacks | DER: 97.05 -> 97.97 | Improved |
| use D3 instead of FT-SAM | PreAct-ResNet18 backdoored model · CIFAR-10 avg of 7 attacks | ACC: 92.46 -> 91.93 | Hurt |
| extend weights beyond fine-tuned point (1<t<2) along trajectory (ablation) | vanilla fine-tuned model · CIFAR-10 four attacks | backdoor loss / ASR | Improved |
| increase lambda from 1 to 40 (ablation) | D3 · CIFAR-10 avg of BadNets/Blended/WaNet | ASR: 0.16 -> 9.40 | Hurt |
| increase lambda from 1 to 40 (ablation) | D3 · CIFAR-10 avg of BadNets/Blended/WaNet | ACC: 90.02 -> 92.49 | Improved |
| raise epsilon from 0 to 0.5 (ablation) | D3 · CIFAR-10 avg of BadNets/Blended/WaNet | ASR: 0.84 -> 0.18 | Improved |
| raise epsilon from 0 to 0.5 (ablation) | D3 · CIFAR-10 avg of BadNets/Blended/WaNet | ACC: 92.29 -> 91.88 | Hurt |
| increase reserved data from 1% to 10% (ablation) | D3 · CIFAR-10 BadNets | ASR: 2.31 -> 0.44 | Improved |
| increase reserved data from 1% to 10% (ablation) | D3 · CIFAR-10 BadNets | ACC: 88.57 -> 90.97 | Improved |
| increase poisoning ratio from 1% to 50% (ablation) | D3 · CIFAR-10 BadNets | ACC: 92.18 -> 86.90 | Hurt |
Closest papers by meaning
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.90)
- Mitigating Backdoor Attack by Injecting Proactive Defensive Backdoor (NeurIPS 2024 · similarity 0.90)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.90)
- Backdoor Secrets Unveiled: Identifying Backdoor Data with Optimized Scaled Prediction Consistency (ICLR 2024 · similarity 0.90)
- Uncovering, Explaining, and Mitigating the Superficial Safety of Backdoor Defense (NeurIPS 2024 · similarity 0.89)
- SampDetox: Black-box Backdoor Defense via Perturbation-based Sample Detoxification (NeurIPS 2024 · 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.
