Research map / Adversarial robustness and security / Backdoor and poisoning
Backdoor Defense via Adaptively Splitting Poisoned Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Adaptively split training data into labeled clean and unlabeled polluted pools, then train with semi-supervision | ResNet-18 · CIFAR-10, six-attack average | ASR: 17.0 -> 1.1 | Helped |
| Add the meta-split stage to recover model-dependent clean hard samples (ablation) | ASD · CIFAR-10, clean accuracy | ACC: about 80% | Helped |
| Add stage 2 class-agnostic loss-guided split (ablation) | ASD · CIFAR-10, four attacks | ACC | Helped |
| Add stage 1 warming-up with progressively growing seed pool (ablation) | ASD · CIFAR-10, four attacks | ACC | Helped |
| Use class-agnostic rather than class-aware splitting in stage 1 (ablation) | ASD · CIFAR-10, four attacks | ACC | Hurt |
| Use class-aware rather than class-agnostic splitting in stage 2 (ablation) | ASD · CIFAR-10, BadNets | ASR: 20% | Hurt |
| Build the clean pool faster with smaller t and larger n (ablation) | ASD stage 1 · CIFAR-10, especially WaNet and CLB | ACC | Hurt |
| Replace MixMatch with UDA (ablation) | ASD · CIFAR-10, BadNets | ACC: 93.4 -> 92.6 | No clear effect |
| Replace MixMatch with ReMixMatch (ablation) | ASD · CIFAR-10, BadNets | ACC: 93.4 -> 91.6 | No clear effect |
| Use 10 rather than 100 clean seed samples (ablation) | ASD · CIFAR-10, Blend | ASR: 1.6 -> 99.1 | Hurt |
| Use 50 rather than 100 clean seed samples (ablation) | ASD · CIFAR-10, Blend | ASR: 1.6 -> 10.4 | Hurt |
| Allow poisoned samples into the seed set (ablation) | ASD · CIFAR-10, Blend | ASR: 1.6 -> 99.9 | Mixed |
Closest papers by meaning
- Bi-perspective Splitting Defense: Achieving Clean-Seed-Free Backdoor Security (ICML 2025 · similarity 0.90)
- Protecting against simultaneous data poisoning attacks (ICLR 2025 · similarity 0.90)
- Chronic Poisoning: Backdoor Attack against Split Learning (AAAI 2024 · similarity 0.89)
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.89)
- Backdoor Defense via Test-Time Detecting and Repairing (CVPR 2024 · similarity 0.89)
- Efficient Backdoor Attacks for Deep Neural Networks in Real-world Scenarios (ICLR 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.