Robust CurveMoE enhances adversarial defense for neural networks

Researchers have developed Robust CurveMoE, a novel mixture-of-experts framework designed to enhance adversarial defense in neural networks. This approach connects models specialized for different perturbation norms through a low-loss path, allowing them to leverage complementary robustness profiles. The framework efficiently derives experts from specific curve locations and selectively applies them to influential layers, while sharing parameters across routing paths. Experiments on CIFAR-100 and ImageNet-100 datasets demonstrated that Robust CurveMoE significantly improves clean, norm-specific, and Union accuracy compared to existing methods. AI

IMPACT Introduces a more efficient method for adversarial defense in neural networks, potentially improving model security against various perturbations.

RANK_REASON The cluster contains a research paper detailing a new model architecture and its performance on benchmarks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• CIFAR-100
• ImageNet-100
• mixture of experts
• Robust CurveMoE
• vision transformer
• WideResNet

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →