Research map / Adversarial robustness and security / Backdoor and poisoning
Defending against Backdoor Attacks via Module Switching
Weijun Li, Ansh Arora, Xuanli He, Mark Dras, Qiongkai Xu
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 |
|---|---|---|---|
| switch modules between two backdoored models instead of averaging weights | RoBERTa-large · SST-2, BadNet + InsertSent | ASR avg: 31.9 -> 22.0 | Improved |
| use module switching instead of WAG on stealthier attack pair | RoBERTa-large · SST-2, BadNet + LWS | ASR avg: 62.2 -> 40.4 | Improved |
| use module switching when merging benign with backdoored model | RoBERTa-large · SST-2, Benign + BadNet | ASR: 39.3 -> 12.2 | Improved |
| use module switching instead of WAG on vision pairs | ViT · CIFAR-10, BadNet + PhysicalBA | ASR avg: 39.6 -> 18.5 | Improved |
| use module switching instead of WAG/TIES on BadNet + WaNet | ViT · CIFAR-10, BadNet + WaNet | ASR avg: WAG 12.2, TIES 11.3, MSD 11.4 | No clear effect |
| use module switching when merging benign with PhysicalBA model | ViT · CIFAR-10, Benign + PhysicalBA | ASR: 10.1 vs 10.1 (WAG) | No clear effect |
| module switching vs WAG on clean accuracy | RoBERTa-large · SST-2, BadNet + InsertSent | CACC: 96.3 -> 96.2 | No clear effect |
| use search without early stopping (higher fitness score) (ablation) | MSD strategy · SST-2 | ASR: -27.2% ASR | Improved |
| add each of the first three heuristic rules (intra-layer, consecutive, residual) to the fitness (ablation) | MSD · SST-2 | ASR | Improved |
| use MSD with three backdoored models instead of WAG | RoBERTa-large · three-model fusion | ASR avg: below 20% | Improved |
| reduce clean validation data from 50 to 20 samples per class (ablation) | MSD selection · SST-2 | ASR of selected candidate | No clear effect |
| apply MSD to CNNs (ResNet-18/50) instead of WAG | ResNet · CIFAR-10 | ASR | Mixed |
Closest papers by meaning
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.90)
- REFINE: Inversion-Free Backdoor Defense via Model Reprogramming (ICLR 2025 · similarity 0.89)
- Fusing Pruned and Backdoored Models: Optimal Transport-based Data-free Backdoor Mitigation (AAAI 2025 · similarity 0.89)
- Mitigating the Backdoor Effect for Multi-Task Model Merging via Safety-Aware Subspace (ICLR 2025 · similarity 0.89)
- BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input Detection (ICLR 2024 · similarity 0.87)
- Backdoor Mitigation via Invertible Pruning Masks (NeurIPS 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.
