Computational Safety for Generative AI: A Hypothesis Testing Perspective

The paper presents a mathematical framework for computational safety in generative AI, focusing on large language models and text-to-image diffusion models. It formalizes safety challenges as hypothesis testing problems, utilizing sensitivity analysis and loss landscape analysis to identify malicious prompts, and applying statistical signal processing to detect AI-generated content. This framework is crucial for practitioners as it provides a systematic approach to developing safety measures that can differentiate responsible AI applications in a landscape where model performance is nearing saturation.