Watch a 45-minute lecture from Miranda Christ of Columbia University exploring the development of pseudorandom error-correcting codes (PRCs) and their application in watermarking AI-generated content. Dive into a novel framework that addresses the growing need to distinguish between human and AI-generated content through hidden pattern embedding. Learn about the cryptographic properties of PRCs, including their ability to generate pseudorandom codewords that remain undetectable to efficient adversaries. Understand the construction of PRCs using subexponential hardness of Learning Parity with Noise (LPN), and discover how these codes maintain robustness against constant rate substitutions and random deletions. Explore practical applications in Large Language Model watermarking, examining both quality preservation and robustness guarantees in the context of alignment, trust, and copyright protection for AI-generated content.
Overview
Syllabus
Pseudorandom Error-Correcting Codes with Applications to Watermarking Generative AI
Taught by
Simons Institute