Link: http://arxiv.org/abs/2502.12445v1
PDF Link: http://arxiv.org/pdf/2502.12445v1
Summary: AI safety is a rapidly growing area of research that seeks to prevent theharm and misuse of frontier AI technology, particularly with respect togenerative AI (GenAI) tools that are capable of creating realistic andhigh-quality content through text prompts.
Examples of such tools include largelanguage models (LLMs) and text-to-image (T2I) diffusion models.
As theperformance of various leading GenAI models approaches saturation due tosimilar training data sources and neural network architecture designs, thedevelopment of reliable safety guardrails has become a key differentiator forresponsibility and sustainability.
This paper presents a formalization of theconcept of computational safety, which is a mathematical framework that enablesthe quantitative assessment, formulation, and study of safety challenges inGenAI through the lens of signal processing theory and methods.
In particular,we explore two exemplary categories of computational safety challenges in GenAIthat can be formulated as hypothesis testing problems.
For the safety of modelinput, we show how sensitivity analysis and loss landscape analysis can be usedto detect malicious prompts with jailbreak attempts.
For the safety of modeloutput, we elucidate how statistical signal processing and adversarial learningcan be used to detect AI-generated content.
Finally, we discuss key openresearch challenges, opportunities, and the essential role of signal processingin computational AI safety.
Published on arXiv on: 2025-02-18T02:26:50Z