Research map / Adversarial robustness and security / Backdoor and poisoning
CatchBackdoor: Backdoor Detection via Critical Trojan Neural Path Fuzzing
Haibo Jin, Ruoxi Chen, Jinyin Chen, Haibin Zheng, Yang Zhang, Haohan Wang
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 |
|---|---|---|---|
| detect trojans via fuzzed critical trojan path instead of trigger-size-dependent reverse engineering | CatchBackdoor · CIFAR-10 AlexNet, BadNets large triggers | LCR: K-Arm N at 18x18 vs CatchBackdoor 72.8% at 18x18 | Improved |
| CatchBackdoor vs baselines on stealthy attacks (blending, defense adaptive) | CatchBackdoor · MNIST LeNet-5, Poison frogs | LCR: 84.4% -> 93.0% | Improved |
| CatchBackdoor on clean-label attacks BP and Sleeper Agent versus K-Arm | CatchBackdoor · MNIST LeNet-4 and CIFAR-10 AlexNet | LCR: BP LeNet-4: 85.0% (K-Arm) vs 83.2% | Hurt |
| increase model depth and dataset scale | CatchBackdoor · a-ImageNet VGG19 | LCR: LCR around 76% on average | Hurt |
| start fuzzing from random noise instead of benign examples (ablation) | CatchBackdoor · MNIST, CIFAR-10, a-ImageNet | LCR: around 73% on average | No clear effect |
| adaptive attacks S-AA and I-AA against CatchBackdoor | CatchBackdoor · adaptive attacks | LCR: LCR over 85% | No clear effect |
| increase top-k neurons k1 and k2 (ablation) | CatchBackdoor · VGG19 a-ImageNet | LCR | Mixed |
| use single neuron (k=1) rather than multi-neuron path against ABE (ablation) | CatchBackdoor · VGG19 a-ImageNet, ABE | LCR | Hurt |
| replace top-k neural path of original class with target class contributions (ablation) | LeNet-5 on MNIST · MNIST target class 1 | consistency rate: 96.6% | Improved |
| LCR threshold between 0.4 and 0.6 (ablation) | CatchBackdoor · 100 benign and trojaned models | AUC: AUC larger than 0.9 | Improved |
Closest papers by meaning
- Trigger Hunting with a Topological Prior for Trojan Detection (ICLR 2022 · similarity 0.91)
- Practical Detection of Trojan Neural Networks: Data-Limited and Data-Free Cases (ECCV 2020 · similarity 0.90)
- Rethinking the Reverse-engineering of Trojan Triggers (NeurIPS 2022 · similarity 0.90)
- BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input Detection (ICLR 2024 · similarity 0.87)
- From Toxic to Trustworthy: Using Self-Distillation and Semi-supervised Methods to Refine Neural Networks (AAAI 2024 · similarity 0.87)
- Scanning Trojaned Models Using Out-of-Distribution Samples (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.
