Research map / Adversarial robustness and security / Attacks on text and graphs
Reinforced Embodied Active Defense: Exploiting Adaptive Interaction for Robust Visual Perception in Adversarial 3D Environments
Xiao Hui Yang, Lingxuan Wu, Lizhong Wang, Chengyang Ying, 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 |
|---|---|---|---|
| add random camera movement (multi-view ensemble) (ablation) | undefended face recognition · CelebA-3D FR impersonation, 3DAdv | attack success rate (%): 98.00 -> 76.77 | Mixed |
| add recurrent perception model (ablation) | passive face recognition · CelebA-3D FR impersonation, 3DAdv | attack success rate (%): 76.77 -> 30.77 | Improved |
| add learned policy model (ablation) | EAD perception-only · CelebA-3D FR impersonation, 3DAdv | attack success rate (%): 30.77 -> 11.34 | Improved |
| add multi-step accumulative interaction with RL (ablation) | EAD · CelebA-3D FR impersonation, adaptive | attack success rate (%): 15.63 -> 6.38 | Improved |
| add multi-step accumulative interaction with RL (ablation) | EAD · CelebA-3D FR impersonation, 3DAdv | attack success rate (%): 11.34 -> 12.77 | Hurt |
| generate OAPA patches with PGD instead of FGSM (ablation) | Rein-EAD · CelebA-3D FR impersonation, 3DAdv | attack success rate (%): 13.18 -> 7.37 | Improved |
| add OAPA (PGD) to multi-step Rein-EAD (ablation) | Rein-EAD · CelebA-3D FR impersonation, 3DAdv | attack success rate (%): 12.77 -> 7.37 | Improved |
| use uncertainty-oriented reward instead of entropy deduction reward (ablation) | Rein-EAD · CelebA-3D FR, 3DAdv | attack success rate (%): 11.42 -> 4.21 | Improved |
| use uncertainty-oriented reward instead of binary outcome reward (ablation) | Rein-EAD · CelebA-3D FR, 3DAdv | attack success rate (%): 10.86 -> 4.21 | Improved |
| replace EAD with Rein-EAD | EAD · CelebA-3D FR impersonation, worst-case | attack success rate (%): 9.38 -> 7.37 | Improved |
| replace EAD with Rein-EAD | EAD · CelebA-3D FR dodging, worst-case | attack success rate (%): 22.11 -> 14.43 | Improved |
| replace EAD with Rein-EAD | EAD · CelebA-3D FR standard accuracy | accuracy (%): 90.45 -> 89.03 | Hurt |
Closest papers by meaning
- Embodied Active Defense: Leveraging Recurrent Feedback to Counter Adversarial Patches (ICLR 2024 · similarity 0.95)
- Adversarial Robust Safeguard for Evading Deep Facial Manipulation (AAAI 2024 · similarity 0.86)
- AdvEDM: Fine-grained Adversarial Attack against VLM-based Embodied Agents (NeurIPS 2025 · similarity 0.85)
- AdvAD: Exploring Non-Parametric Diffusion for Imperceptible Adversarial Attacks (NeurIPS 2024 · similarity 0.85)
- HASARD: A Benchmark for Vision-Based Safe Reinforcement Learning in Embodied Agents (ICLR 2025 · similarity 0.85)
- A Unified, Resilient, and Explainable Adversarial Patch Detector (CVPR 2025 · similarity 0.85)
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.
