Watermarking in Generative AI: Opportunities and Threats
From Google TechTalks · 59:07
Show by Google TechTalks
A Google TechTalk, 2025-06-18, presented by Nikola JovanovićPrivacy in ML Seminar. ABSTRACT: As generative AI becomes more capable and increasingly integrated into everyday tools and platforms, watermarking has emerged as a promising technique for content attribution, provenance tracking, and misuse detection. This talk presents recent advances and challenges in watermarking, focusing on autoregressive generative models. First, we expose vulnerabilities in current LLM watermarking schemes by introducing a novel class of adversaries capable of cheaply mounting both spoofing and removal attacks. Second, we shift to defense, demonstrating how LLM watermarks can be repurposed to enable data owners to provably detect unauthorized use of their content in retrieval-augmented generation (RAG) systems. Finally, we extend token-level watermarking to autoregressive image generation models, a paradigm that has recently gained traction, by identifying and addressing key modality-specific challenges, enabling a theoretically principled, strong, and robust watermark.Uploaded 26-01-26
More episodes of Google TechTalks
More episodes
- Threat Models for Memorization: Privacy, Copyright, and Everything In-Between
- Technology Assisted Reading Acquisition: Children Acquiring Literacy Naturally
- sex on the internet, the realities of porn, sexual privacy,
- Privacy Auditing of Large Language Models
- The Limits and Possibilities of One Run Auditing
- PhotoTechEDU Day 29: Photographing VR Panoramas
- Privacy Ripple Effects from Adding or Removing Personal Information in Language Model Training
- Inside VMware Fusion
- Java Posse Episode 100 (Live Special)
- Cascading Adversarial Bias from Injection to Distillation in Language Models
PodSearch is a free, independent podcast search engine. Read about PodSearch, the privacy policy or contact us.