Research map / Adversarial robustness and security / Black-box and transfer attacks
PP-Mark: Provable and Publicly Verifiable Watermarking for Generative AI
HANKYEONG KO, Taehoon Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PP-Mark, a provenance framework that ties a public, lightweight statistical watermark detector to a zero-knowledge proof of the embedding process so that generated content can be publicly verified without enabling forgery or adversarial removal.
Problem
Publicly verifiable provenance for generative AI images is needed, but existing watermarks become vulnerable to forgeries and adversarial optimization attacks once their detectors are made public.
Approach
Anchors lightweight statistical watermark detection to a zero-knowledge proof of the embedding process, so the public can verify that content was watermarked correctly without the detector being exploitable for forgery or removal.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ZKP receipt verification (accept mode) on top of score (ablation) | PP-Mark · White-box PGD forgery | FAR@1%FPR: 0.76 -> 0.00 | Improved |
| add ZKP receipt verification on top of signature+score (ablation) | Sig+Score · Forged binding attack | FAR: 1.00 -> 0.00 | Improved |
| bind watermark to secret key via context-bound latent embedding (score only) (ablation) | PP-Mark (score) · Forged binding attack | FAR: 1.00 -> <=0.01 | Improved |
| add score check to signature-only verification (ablation) | Sig-only · Compliance bypass attack | FAR: 1.00 -> 0.00 | Improved |
| add score check to signature-only verification (ablation) | Sig-only · Forged binding attack | FAR: 1.00 -> 1.00 | No clear effect |
| use PP-Mark score detector instead of other detectors | PP-Mark (score) · Imprint forgery transfer, Tree-Ring source | FAR@1%FPR: 0.0400 -> 0.0033 | Improved |
| use PP-Mark score detector instead of WIND under imprint forgery from GS | PP-Mark (score) · Imprint forgery transfer, GS source | FAR@1%FPR: 0.0100 -> 0.0067 | No clear effect |
| increase watermark strength alpha from 4.0 to 4.5 (ablation) | PP-Mark · SD2.1 image quality | BRISQUE p95: 34.9 -> 40.5 | Hurt |
| increase number of sampled positions N from 200 to 1000 (ablation) | PP-Mark · Score variance | score std across images: 0.1409 -> 0.1301 | Improved |
| increase N from 1000 to 1500 (ablation) | PP-Mark · Score variance | score std across images: 3.0% additional reduction | No clear effect |
| increase N from 200 to 1000 (SP1 v5) (ablation) | PP-Mark proof · SP1 proving cost | prove time (s): 32.309 -> 48.572 | Hurt |
| switch SP1 v5 to v6.3.1 circuit (ablation) | PP-Mark proof · N=1000 proving | prove time (s): 48.572 -> 17.0 | Improved |
Closest papers by meaning
- WMCopier: Forging Invisible Watermarks on Arbitrary Images (NeurIPS 2025 · similarity 0.89)
- Watermarks in the Sand: Impossibility of Strong Watermarking for Language Models (ICML 2024 · similarity 0.88)
- An Undetectable Watermark for Generative Image Models (ICLR 2025 · similarity 0.87)
- Break the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models (ICLR 2026 · similarity 0.87)
- No Free Lunch in LLM Watermarking: Trade-offs in Watermarking Design Choices (NeurIPS 2024 · similarity 0.87)
- Disentangled Style Domain for Implicit $z$-Watermark Towards Copyright Protection (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
