Research map / Adversarial robustness and security / Certified and empirical robustness
On the Adversarial Robustness of Mixture of Experts
Joan Puigcerver, Rodolphe Jenatton, Carlos Riquelme Ruiz, Pranjal Awasthi, Srinadh Bhojanapalli
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 dense ViT-B/32 with sparse V-MoE-B/32 | ViT-B/32 · ImageNet adversarial attacks | classification error | Improved |
| Replace dense ViT-B/32 with sparse V-MoE-B/32 | ViT-B/32 · JFT-300M adversarial attacks | false discovery rate | Improved |
| PGD adversarial fine-tuning | ViT-B/32 and V-MoE-B/32 · ImageNet | classification error: 51.7% and 49.8% | Hurt |
| Increase the total experts from 2 to 16 (ablation) | V-MoE-B/32 · JFT-300M adversarial attacks | false discovery rate | Improved |
| Increase the total experts from 16 to 32 (ablation) | V-MoE-B/32 · JFT-300M adversarial attacks | false discovery rate | No clear effect |
| Use more than 2 total experts (ablation) | V-MoE-B/32 · ImageNet adversarial attacks | classification error | No clear effect |
| Train the V-MoE model with PGD adversarial fine-tuning (ablation) | V-MoE-B/32 · ImageNet adversarial attacks | routing changes | Improved |
| Include the router auxiliary losses in the attack objective (ablation) | V-MoE-B/32 · JFT-300M adversarial attacks | false discovery rate | No clear effect |
| Use AutoPGD instead of PGD | ViT-B/32 · JFT-300M at ℓ∞ = 10⁻³ | false discovery rate: 0.744 -> 0.757 | Hurt |
| Use AutoPGD instead of PGD | V-MoE-B/32 · ImageNet adversarial attacks | routing changes: near 0.2 -> around 0.5 | Hurt |
Closest papers by meaning
- Optimizing Robustness and Accuracy in Mixture of Experts: A Dual-Model Approach (ICML 2025 · similarity 0.93)
- Robust Mixture-of-Expert Training for Convolutional Neural Networks (ICCV 2023 · similarity 0.91)
- Robustness of Mixtures of Experts to Feature Noise (ICML 2026 · similarity 0.90)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.88)
- On the Expressive Power of Mixture-of-Experts for Structured Complex Tasks (NeurIPS 2025 · similarity 0.87)
- A Recipe for Improved Certifiable Robustness (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.
