Research map / Adversarial robustness and security / Certified and empirical robustness
Exposing and Defending the Achilles' Heel of Video Mixture-of-Experts
Songping Wang, Qinglong Liu, Yueming Lyu, Ning Li, Ziwen He, Caifeng Shan
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 joint router-and-expert attack (J-TLGA) instead of router-only TLGA-R | Video MoE · UCF-101 eps=8/255 | robust accuracy (%): TLGA-R 18.13 -> J-TLGA 4.95 (stronger attack lowers accuracy) | Improved |
| add Lipschitz regularity and temporal adaptive step to PGD (TLA-M vs PGD) | Video MoE attack · UCF-101 eps=8/255 | robust accuracy (%): 22.09 -> 15.05 (lower is stronger attack) | Improved |
| add Lipschitz term to router attack gradient (TLGA-R vs PGD-R) | Video MoE attack · UCF-101 eps=8/255 | robust accuracy (%): 42.20 -> 18.13 (lower is stronger attack) | Improved |
| add temporal adaptive step to expert attack (TLA-E vs PGD-E) | Video MoE attack · UCF-101 eps=8/255 | robust accuracy (%): 31.98 -> 28.11 (lower is stronger attack) | Improved |
| add joint component adversarial training (J-TLAT) over TLAT | AT-M MoE · UCF-101 PGD eps=8/255 | robust accuracy (%): 30.22 -> 36.37 | Improved |
| add joint component adversarial training (J-TLAT) over TLAT | AT-M MoE · UCF-101 J-TLGA eps=8/255 | robust accuracy (%): 9.78 -> 33.96 | Improved |
| add J-TLAT over TLAT | AT-M MoE · UCF-101 TLGA-R eps=8/255 | robust accuracy (%): 26.37 -> 51.87 | Improved |
| add J-TLAT over TLAT | MoE · UCF-101 router Lipschitz constant | Lips-R: 3.500 -> 0.823 | Improved |
| use J-TLAT instead of AAT-M | MoE · UCF-101 J-TLGA eps=8/255 | robust accuracy (%): 0.00 -> 33.96 | Improved |
| use TLA as inner attack instead of PGD in J-TLAT (ablation) | J-TLAT · UCF-101 PGD eval | robust accuracy (%): 33.63 (a 5.5% gain over conventional PGD) | Improved |
| use TLA as inner attack in J-TLAT (ablation) | J-TLAT · UCF-101 J-TLGA eval | robust accuracy (%): 30.33 | Improved |
| use J-TLAT regularization instead of JS, COS, KL, NR, GR (ablation) | J-TLAT · UCF-101 FGSM/PGD/J-TLGA | robust accuracy (%): 36.59, 33.63, 30.33 | Improved |
Closest papers by meaning
- Optimizing Robustness and Accuracy in Mixture of Experts: A Dual-Model Approach (ICML 2025 · similarity 0.88)
- From Sparse to Soft Mixtures of Experts (ICLR 2024 · similarity 0.85)
- On the Expressive Power of Mixture-of-Experts for Structured Complex Tasks (NeurIPS 2025 · similarity 0.84)
- MoDE: A Mixture-of-Experts Model with Mutual Distillation among the Experts (AAAI 2024 · similarity 0.84)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.84)
- A Recipe for Improved Certifiable Robustness (ICLR 2024 · similarity 0.84)
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.
