Research map / Adversarial robustness and security / Backdoor and poisoning
Prototype Guided Backdoor Defense via Activation Space Manipulation
Venkat Adithya Amula, Sunayana Samavedam, Saurabh Saini, Avani Gupta, P J Narayanan
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 PGBD sanitization loss over fine-tuning/baseline | preact-ResNet18 poisoned model · mean over datasets, Badnet | ASR: 88.67 -> 4.81 | Improved |
| add PGBD sanitization loss | preact-ResNet18 poisoned model · mean over datasets, Trojan | ASR: 99.64 -> 6.99 | Improved |
| use PGBD instead of MCL | preact-ResNet18 · mean over datasets, Badnet | CA: 64.96 -> 75.17 | Improved |
| use PGBD instead of FT-SAM | ROF semantic attacks · ROF | average DEM: 0.84 -> 0.93 | Improved |
| use PGBD instead of MCL under adaptive Wanet* | CIFAR10 · adaptive attack Wanet* | DEM: 0.37 -> 0.94 | Improved |
| use synthesized-trigger PAV (ST-PGBD) instead of target prototype PAV (ablation) | PGBD · CIFAR10 Badnet all-to-all | DEM: 0.99 -> 0.68 | Hurt |
| use synthesized-trigger PAV (ST-PGBD) instead of target prototype PAV (ablation) | PGBD · CIFAR10 semantic attacks | DEM: 0.98 -> 0.80 (Tattoo) | Hurt |
| remove target knowledge, cycle all classes (NT-PGBD) (ablation) | PGBD · CIFAR10 | DEM: 0.95 -> 0.93 (Trojan) | No clear effect |
| add large-model (DINOv1) mapping (ablation) | PGBD · CIFAR10 | CA | Improved |
| add large-model mapping (ablation) | PGBD · CIFAR10 | ASR reduction | Hurt |
| set lambda to 10 (ablation) | PGBD · CIFAR10 | CA/ASR trade-off | Mixed |
| reduce clean data Ds size (ablation) | PGBD · defense-time ablation | CA/ASR | No clear effect |
Closest papers by meaning
- Breaking the False Sense of Security in Backdoor Defense through Re-Activation Attack (NeurIPS 2024 · similarity 0.89)
- DEFEAT: Deep Hidden Feature Backdoor Attacks by Imperceptible Perturbation and Latent Representation Constraints (CVPR 2022 · similarity 0.88)
- Hidden Trigger Backdoor Attacks (AAAI 2020 · similarity 0.88)
- REFINE: Inversion-Free Backdoor Defense via Model Reprogramming (ICLR 2025 · similarity 0.87)
- Mitigating Backdoor Attack by Injecting Proactive Defensive Backdoor (NeurIPS 2024 · similarity 0.87)
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · 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.
