Entity · paper

Watermark Forensics for Generative Models: An Information-Theoretic Perspective

paperactiveprovisionalwatermark-forensics-for-generative-models-an-information-theoretic-perspective-f33bf293·1 events·first seen 6d ago

Aliases: Watermark Forensics for Generative Models: An Information-Theoretic Perspective

Co-occurring entities

Qwen2.5 GPT-2 Pythia-410M

More like this (12)

Selective Disclosure Watermarking for Large Language Models AI Watermark Evidence Fails Forensic Readiness: An Empirical Evaluation Watermarking generative models Unigram watermarking diffusion-based generative models generative language modeling causal generative modeling Simple Domain Generalization for Strong Pixel-Level Image Tampering Detection in Modern VLMs NAMESAKES: Probing Identity Memorization in Text-to-Image Models Generative Adversarial Networks Variational Information Bottleneck

Recent events (1)

6arXiv · cs.LG·6d ago·source ↗

Information-theoretic framework establishes tight sample complexity laws for watermark forensics in generative models

A new arXiv preprint develops an information-theoretic framework for watermark forensics in generative model outputs, organizing detection, attribution, payload extraction, and localization into a 'forensic ladder' with precise sample complexity bounds. The main theorem establishes the first tight entropy-rate law for multi-user attribution: attributing text to one of N users costs Θ(log N/h) tokens under statistically distortion-free schemes, with a matching converse. The paper also identifies two fundamental gaps — a window where text is provably machine-made but unattributable, and a footprint-resolution uncertainty principle — validated experimentally on GPT-2, Pythia-410M, and Qwen2.5.

Evaluation and Benchmarking AI Safety Research Watermark Forensics for Generative Models: An Information-Theoretic Perspective Qwen2.5 GPT-2 +1 more