Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore the cutting-edge field of watermarking in generative AI through this comprehensive technical presentation that examines both defensive applications and security vulnerabilities. Delve into recent advances and challenges in watermarking techniques specifically designed for autoregressive generative models, including large language models and image generation systems. Discover how current LLM watermarking schemes face significant security threats through a novel class of adversaries capable of executing both spoofing and removal attacks at low cost. Learn how these same watermarking technologies can be strategically repurposed as defensive tools, enabling data owners to provably detect unauthorized use of their content within retrieval-augmented generation (RAG) systems. Examine the extension of token-level watermarking principles to autoregressive image generation models, understanding the unique modality-specific challenges that arise when transitioning from text to visual content. Gain insights into the theoretical foundations that enable the development of strong, robust watermarking solutions that maintain effectiveness across different generative AI paradigms. Understand the dual nature of watermarking technology as both a protective measure for content attribution and provenance tracking, while also recognizing the sophisticated attack vectors that threaten current implementations in the rapidly evolving landscape of generative artificial intelligence.
