Megha Srivastava (Stanford)- New Challenges of Trust with Large-Scale AI Systems

Abstract: Today’s large-scale AI systems, trained with > 200 billion parameters over massive datasets, create new challenges of trust as users have increasingly less control over all aspects of model development.
I will first do a deep dive on the challenge of auditing model training service providers, who currently fine-tune models on behalf of resource-poor clients for a fee without any guarantee of correctness. I will show how prior solutions to this “verifiable training” problem are non-robust due to hardware non-determinism, which we address via a threshold-based rounding scheme over intermediate computations during model training. I will then show how users of modern code-generation models may introduce accidental security vulnerabilities due to misplaced confidence. Finally, I will conclude by discussing ongoing work on the current limitations of methods that seek to establish trust via content provenance (e.g. watermarking, C2PA).