Research map / Adversarial robustness and security / Backdoor and poisoning
Towards Stable Backdoor Purification through Feature Shift Tuning
Rui Min, Zeyu Qin, Li Shen, Minhao Cheng
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 Feature Shift Tuning instead of the strongest compared defense | Backdoor purification on CIFAR-10, GTSRB, and Tiny-ImageNet · CIFAR-10, GTSRB, Tiny-ImageNet | ASR: 34.04% and 26.91% | Improved |
| Use Feature Shift Tuning instead of ANP or I-BAU | Backdoor purification on CIFAR-10 and GTSRB · CIFAR-10 and GTSRB | ASR: 11.34% and 32.16% | Improved |
| Add feature-extractor tuning with a reinitialized, frozen classifier (FE-tuning) rather than LP (ablation) | Backdoor purification · CIFAR-10, BadNet, Blended, SSBA, and LC at 1% and 0.5% poisoning | ASR: 90.91 -> 15.11 | Improved |
| Add whole-model tuning with a reinitialized classifier (FT-init) rather than vanilla FT (ablation) | Backdoor purification · CIFAR-10, BadNet, Blended, SSBA, and LC at 1% and 0.5% poisoning | ASR: 74.36 -> 52.60 | Improved |
| Use FST rather than FE-tuning (ablation) | Backdoor purification · Tiny-ImageNet | C-Acc: 8.1% | Improved |
| Add the classifier weight-deviation penalty and projection constraint (FST) rather than FT-init (ablation) | Backdoor purification · CIFAR-10 and GTSRB | ASR | Improved |
| Add the classifier weight-deviation penalty and projection constraint (FST) rather than FT-init (ablation) | Backdoor purification · CIFAR-100 and Tiny-ImageNet | ASR: 32.31% | Mixed |
| Add the projection constraint (ablation) | Feature Shift Tuning · CIFAR-10, BadNet and Blended | convergence stability | Improved |
| Increase α from 0 to 0.1, then continue increasing it (ablation) | Feature Shift Tuning · CIFAR-10, BadNet and Blended | ASR: ASR < 5%; ASR < 4%; accuracy degradation < 1.5% | Improved |
| Reduce the tuning set to 0.1% of training data (ablation) | Feature Shift Tuning · CIFAR-10 | ASR: less than 2% | No clear effect |
| Tune for only a few epochs with FST rather than other tuning defenses (ablation) | Backdoor purification · CIFAR-10, 1% poisoning | ASR: below 5% | Improved |
| Extend FST's weight-deviation penalty from the last classifier layer to all linear classifier layers (ablation) | FST on VGG19-BN · CIFAR-10 | ASR | Improved |
Closest papers by meaning
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.92)
- Uncovering, Explaining, and Mitigating the Superficial Safety of Backdoor Defense (NeurIPS 2024 · similarity 0.91)
- BAN: Detecting Backdoors Activated by Adversarial Neuron Noise (NeurIPS 2024 · similarity 0.89)
- REFINE: Inversion-Free Backdoor Defense via Model Reprogramming (ICLR 2025 · similarity 0.89)
- From Toxic to Trustworthy: Using Self-Distillation and Semi-supervised Methods to Refine Neural Networks (AAAI 2024 · similarity 0.89)
- BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input Detection (ICLR 2024 · 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.
