Research map / Adversarial robustness and security / Certified and empirical robustness
Robust Mixture-of-Expert Training for Convolutional Neural Networks
Yihua Zhang, Ruisi Cai, Tianlong Chen, Guanhua Zhang, Huan Zhang, Pin-Yu Chen, Shiyu Chang, Zhangyang Wang, Sijia Liu
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 |
|---|---|---|---|
| train MoE-CNN with router-expert alternating bi-level AdvMoE instead of standard AT | MoE-CNN ResNet-18 · CIFAR-10 (PGD robust accuracy) | RA-PGD (%): 45.57 -> 51.83 | Improved |
| AdvMoE (r=0.5 pathway) versus dense model with twice the width | WRN-28-10 · CIFAR-10 (PGD robust accuracy) | RA-PGD (%): 51.75 -> 55.73 | Improved |
| AdvMoE versus dense baseline on clean accuracy with ResNet-18 | MoE-CNN ResNet-18 · CIFAR-10 (clean accuracy) | SA (%): 82.99 -> 80.15 | Hurt |
| use MoE-CNN pathway instead of dense model | AdvMoE · CIFAR-10 (inference cost) | GFLOPS: 0.54 -> 0.15 | Improved |
| evaluate with AutoAttack instead of PGD | AdvMoE · CIFAR-10 (AutoAttack) | RA-AA (%): 40.42 -> 45.13 | Improved |
| increase number of experts N to 10 (with r = 0.8) (ablation) | AdvMoE · CIFAR-10 ResNet-18 expert count sweep | RA | Hurt |
| use small per-expert scale r<0.5 with N>=8 (ablation) | AdvMoE · CIFAR-10 ResNet-18 expert count sweep | RA | Hurt |
| increase model scale r from 0.2 to 0.8 (ablation) | AdvMoE · CIFAR-10 | RA | Improved |
| raise training PGD steps from 2 to 10 (ablation) | AdvMoE · CIFAR-10 ResNet-18 | RA (%): 51.83 -> 52.22 | No clear effect |
| apply AdvMoE to MoE-ViT | DeiT-Tiny MoE · ImageNet MoE-ViT | RA (%): 44.63 -> 45.93 | Improved |
Closest papers by meaning
- Optimizing Robustness and Accuracy in Mixture of Experts: A Dual-Model Approach (ICML 2025 · similarity 0.92)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.89)
- From Sparse to Soft Mixtures of Experts (ICLR 2024 · similarity 0.88)
- Dense Backpropagation Improves Training for Sparse Mixture-of-Experts (NeurIPS 2025 · similarity 0.88)
- On the Expressive Power of Mixture-of-Experts for Structured Complex Tasks (NeurIPS 2025 · similarity 0.87)
- Efficient Data Driven Mixture-of-Expert Extraction from Trained Networks (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.
