Research map / Adversarial robustness and security / Backdoor and poisoning
Logit-Margin Repulsion for Backdoor Defense
Zhiguo Yang, Dongsheng Xu, Ruizhi Zhong, Jiacheng Pi, Xingxing Huang, Wenjie Ruan
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 |
|---|---|---|---|
| add directed logit-margin suppression (DSC) to selective CE (ablation) | LMR · CIFAR-10 BadNets, ResNet-18 | ASR: 92.70 -> 69.57 (m1=1), 4.84 (m1=10) | Improved |
| increase DSC margin m1 from 1 to 10 (ablation) | LMR · CIFAR-10 BadNets, ResNet-18 | ASR: 69.57 -> 4.84 | Improved |
| add conditional margin (CM) to SCE+DSC (ablation) | LMR · CIFAR-10 BadNets, ResNet-18 | ASR: 4.84 -> 0.69 | Improved |
| add CM alone to SCE (no DSC) (ablation) | LMR · ViT ImageNet | ASR: 93.09 -> 64.91; full 0.91 | Improved |
| replace CE-only fine-tuning with full LMR losses (ablation) | LMR · ViT ImageNet | ASR: 93.09 -> 0.91 | Improved |
| apply LMR instead of no defense | LMR · CIFAR-10 traditional attacks avg | ASR: 96.80 -> 0.53 | Improved |
| apply LMR instead of no defense | LMR · CIFAR-10 traditional attacks avg | accuracy: 95.55 -> 95.03 | Hurt |
| apply LMR instead of no defense | LMR · ImageNet traditional attacks avg | ASR: 94.18 -> 0.68 | Improved |
| apply LMR to conditional backdoors vs QCB-specialized defenses | LMR · CIFAR-10 conditional attacks avg | ASR: 33.53 (EFRAP), 34.08 (LACPDA) -> 0.72 | Improved |
| apply LMR on PCB model after 50% pruning | LMR · CIFAR-10 PCB | ASR: 99.99 -> 1.54 | Improved |
| apply LMR to 8-bit QCB across architectures | LMR · CIFAR-10 QCB (ResNet-18, VGG-16, MobileNetV2) | ASR: 99.81 -> 0.47 (ResNet-18) | Improved |
| vary alpha in [0.5, 3] (ablation) | LMR · CIFAR-10 | ACC: ACC within [83.7, 94.4] | Mixed |
Closest papers by meaning
- BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input Detection (ICLR 2024 · similarity 0.88)
- Backdoor Mitigation via Invertible Pruning Masks (NeurIPS 2025 · similarity 0.87)
- BackdoorDM: A Comprehensive Benchmark for Backdoor Learning on Diffusion Model (NeurIPS 2025 · similarity 0.87)
- REFINE: Inversion-Free Backdoor Defense via Model Reprogramming (ICLR 2025 · similarity 0.87)
- Inspecting Prediction Confidence for Detecting Black-Box Backdoor Attacks (AAAI 2024 · similarity 0.87)
- Uncovering, Explaining, and Mitigating the Superficial Safety of Backdoor Defense (NeurIPS 2024 · 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.
