Information Hiding — Watermarking and Steganography
Stub. Hiding information inside other content rather than (or in addition to) encrypting it: the information-theoretic formulation of watermarking as communication over a channel with side information (Moulin-O'Sullivan, Chen-Wornell), robustness/capacity tradeoffs (Moulin-Koetter), and the modern application of watermarking outputs of generative language models (Dathathri et al.) to make AI-generated text detectable.
Corresponds to CS/ECE 598DA, Week 13. Watermarking embeds a hidden signal into content (an image, a document, a stream of generated text) that survives ordinary processing of that content and can later be detected or extracted, which is a different goal from encryption — the content itself stays usable and (ideally) visually/statistically unchanged, rather than being scrambled. Moulin-O’Sullivan and Chen-Wornell frame this information-theoretically: the watermarker has side information (the original, unwatermarked content) that an adversary trying to remove the watermark doesn’t, which sets up a capacity/ robustness tradeoff analyzable with the same tools as a communication channel with side information. Dathathri et al. is the direct modern payoff: watermarking large language model outputs by biasing the token- sampling distribution in a way that’s statistically detectable later but doesn’t noticeably degrade text quality — the same idea applied to a generative model’s output stream instead of an image or document.
Outline (TODO — flesh out each)
- Watermarking as communication with side information (Moulin-O’Sullivan)
- Capacity and robustness tradeoffs (Moulin-Koetter)
- Spread-spectrum and quantization-based watermarking (Chen-Wornell)
- Watermarking generative language model outputs (Dathathri et al.)
- Steganography