New framework makes AI model robustness checks cheaper and faster

Researchers have developed a new framework to make model robustness certificates more efficient. This method uses a cheap proxy function to statistically control deviations from expensive, sound certificates like randomized smoothing. By tuning the proxy with robust conformal risk control, the framework guarantees a bounded marginal false-approval rate, meaning the probability that the proxy accepts an input whose prediction is flipped by a perturbation. This approach can reduce certification costs to a single model forward pass while retaining significant certified performance for large architectures such as Vision Transformers and Language Models. AI

IMPACT Enables more cost-effective and practical robustness guarantees for large AI models.

RANK_REASON Academic paper detailing a new technical approach to AI model robustness. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Language Models
• Soroush H. Zargarbashi
• Vision Transformers

• paper
• safety

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