Research map / Adversarial robustness and security / Backdoor and poisoning
Black-Box Detection of Backdoor Attacks With Limited Information and Data
Yinpeng Dong, Xiao Yang, Zhijie Deng, Tianyu Pang, Zihao Xiao, Hang Su, Jun Zhu
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 |
|---|---|---|---|
| replace gradient-based NC with gradient-free B3D | backdoor detection · CIFAR-10 | detection accuracy: 95.0% -> 97.5% | Improved |
| use Bernoulli discrete mask (B3D) instead of relaxed continuous mask (NC) | trigger reverse-engineering · CIFAR-10 normal models | normal models wrongly flagged (of 50): 8/50 -> 2/50 | Improved |
| use synthetic samples instead of clean data (B3D-SS vs B3D) | B3D · ImageNet (10-class subset) | detection accuracy: 96.0% -> 95.5% | No clear effect |
| use synthetic samples instead of clean data (B3D-SS vs B3D) | B3D · CIFAR-10 | detection accuracy: 97.5% -> 97.5% | No clear effect |
| increase trigger size from 1x1 to 3x3 | NC · CIFAR-10 | Case I detections (of 50): 40/50 -> 47/50 -> 49/50 | Improved |
| use B3D reversed-trigger KL score instead of STRIP | triggered-input detection · CIFAR-10 | AUC: 0.9332 -> 0.9958 | Improved |
| use B3D reversed-trigger KL score instead of STRIP | triggered-input detection · GTSRB | AUC: 0.4937 -> 0.9946 | Improved |
| vary number of gradient samples k within 20 to 100 (ablation) | B3D · CIFAR-10 | detection accuracy: k in [20,100] leads to similar results | No clear effect |
| use BigGAN synthetic images instead of random noise (ablation) | B3D-SS · ImageNet (10-class subset) | detection quality | Improved |
| add heuristic rule (mask L1 below one fourth of median) to MAD outlier detection (ablation) | NC and B3D · CIFAR-10 | detection accuracy | Improved |
| add data augmentation to backdoored model training | backdoored ResNet-18 · CIFAR-10 | clean accuracy and ASR: accuracy 94.68%, 94.78%, 95.29%; ASR 99.67%, 99.99%, 100.00% | Improved |
| switch ResNet-18 to VGG-16 | B3D · CIFAR-10 | detection accuracy: 97.5% -> 97.0% | No clear effect |
Closest papers by meaning
- Inspecting Prediction Confidence for Detecting Black-Box Backdoor Attacks (AAAI 2024 · similarity 0.91)
- Adversarial-Inspired Backdoor Defense via Bridging Backdoor and Adversarial Attacks (AAAI 2025 · similarity 0.90)
- Backdoor Defense via Test-Time Detecting and Repairing (CVPR 2024 · similarity 0.89)
- BAN: Detecting Backdoors Activated by Adversarial Neuron Noise (NeurIPS 2024 · similarity 0.89)
- A Theoretical Analysis of Backdoor Poisoning Attacks in Convolutional Neural Networks (ICML 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.
