Research map / Adversarial robustness and security / Backdoor and poisoning
Marksman Backdoor: Backdoor Attacks with Arbitrary Target Class
Khoa D Doan, Yingjie Lao, Ping Li
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 |
|---|---|---|---|
| Use a class-conditional generative trigger and constrained joint optimization instead of repeated fixed trigger patterns | Marksman versus multi-trigger baselines · MNIST, CIFAR10, GTSRB, T-IMNET at 10% poisoning | Attack success rate: MNIST 0.784 -> 1.000; CIFAR10 0.730 -> 1.000; GTSRB 0.802 -> 0.997; T-IMNET 0.376 -> 0.999 | Improved |
| Use Marksman rather than repeated multi-trigger baselines | Marksman · MNIST, CIFAR10, GTSRB, T-IMNET at 10% poisoning | Clean accuracy: MNIST: 0.969 -> 0.983; CIFAR10: 0.934 -> 0.943; GTSRB: 0.951 -> 0.979; T-IMNET: 0.577 -> 0.575 | Mixed |
| Use Marksman rather than repeated multi-trigger baselines | Marksman · MNIST, CIFAR10, GTSRB, T-IMNET at 50% poisoning | Attack success rate: MNIST: 0.909 -> 1.000; CIFAR10: 0.999 -> 1.000; GTSRB: 0.999 -> 0.999; T-IMNET: 0.999 -> 0.999 | Improved |
| Transfer the learned trigger generator to a separately trained classifier | Marksman trigger · CIFAR10 and GTSRB | Attack success rate: CIFAR10: 0.990 -> 0.999; GTSRB: 0.993 -> 0.997 | Improved |
| Set alpha to 1.0, removing classifier poisoning (ablation) | Marksman · CIFAR10 | Attack success rate: ASR drops to 0 | Hurt |
| Lower alpha from its near-optimal value (ablation) | Marksman · CIFAR10 | Clean accuracy | Hurt |
| Vary beta while holding alpha at 0.8 (ablation) | Marksman · CIFAR10 | Clean and attack performance | No clear effect |
| Update the trigger generator less frequently than the classifier | Marksman optimization · Training | Convergence speed | Improved |
| Apply Neural Cleanse to the Marksman model | Marksman · MNIST, CIFAR10, GTSRB, T-IMNET | Backdoor detection | No clear effect |
| Apply STRIP to detect Marksman-triggered inputs | Marksman · MNIST, CIFAR10, GTSRB, T-IMNET | Detection separation | No clear effect |
| Apply Spectral Signature to Marksman samples | Marksman · MNIST, CIFAR10, GTSRB, T-IMNET | Latent-space separation | No clear effect |
| Apply Fine-Pruning to the Marksman model | Marksman · MNIST, CIFAR10, GTSRB, T-IMNET | Attack success rate | No clear effect |
Closest papers by meaning
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.90)
- Architectural Backdoors in Neural Networks (CVPR 2023 · similarity 0.90)
- Hidden Trigger Backdoor Attacks (AAAI 2020 · similarity 0.89)
- Mitigating Backdoor Attack by Injecting Proactive Defensive Backdoor (NeurIPS 2024 · similarity 0.88)
- Data Free Backdoor Attacks (NeurIPS 2024 · similarity 0.88)
- BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input Detection (ICLR 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.
