Research map / Image generation / Autoregressive and latent generation
TrustMark: Robust Watermarking and Watermark Removal for Arbitrary Resolution Images
Tu Bui, Shruti Agarwal, John Collomosse
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GAN loss (ablation) | TrustMark without GAN, FFL, Epost · DIV2K | PSNR (dB): +0.7dB | Improved |
| add 1x1-conv post-process module Epost (ablation) | TrustMark without GAN, FFL, Epost · DIV2K | PSNR (dB): +1.6dB | Improved |
| add focal frequency loss (FFL) (ablation) | TrustMark without GAN, FFL, Epost · DIV2K | PSNR (dB): +1.7dB | Improved |
| add GAN, Epost and FFL together (ablation) | TrustMark · DIV2K | PSNR (dB) and bit accuracy: 36.30 -> 40.20 PSNR, 0.959 -> 0.973 accuracy | Improved |
| combine Epost with GAN or FFL (ablation) | TrustMark · DIV2K | PSNR (dB): boosts PSNR by 3dB | Improved |
| use VGG, GoogleNet, ConvNext or EfficientNet as extractor instead of ResNet family (ablation) | TrustMark extractor · DIV2K | training convergence: training not successful | Hurt |
| add noise simulation during training (no noise vs high severity) (ablation) | TrustMark · DIV2K | PSNR (dB): 53.2dB -> 40.2dB PSNR, bit accuracy reduces by only 3% | Hurt |
| increase alpha_max to 30 (ablation) | TrustMark · DIV2K | PSNR and bit accuracy: PSNR improves by more than 8dB while bit accuracy drops by less than 5% | Mixed |
| increase watermark length from 32 to 200 bits (ablation) | TrustMark · DIV2K | PSNR and bit accuracy: PSNR and bit accuracy both drop by 7.5dB and 11% | Hurt |
| add ReMark removal before re-watermarking | TrustMark · DIV2K | PSNR (dB): 40.20 -> 48.48 PSNR, bit accuracy 0.970 -> 0.553 | Improved |
| add high-severity noise simulation against I-FGSM attack (ablation) | TrustMark · DIV2K | attack iterations: under 50 iterations without noise vs more than 3000 iterations for 32% of images | Improved |
| apply Resolution Scaling across 20% to 100% of 2K resolution (ablation) | TrustMark · DIV2K | PSNR and bit accuracy: PSNR varies only by ±0.02dB, bit accuracy by ±10−4 | No clear effect |
Closest papers by meaning
- An Undetectable Watermark for Generative Image Models (ICLR 2025 · similarity 0.88)
- Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances (ICLR 2025 · similarity 0.88)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.88)
- Invisible Image Watermarks Are Provably Removable Using Generative AI (NeurIPS 2024 · similarity 0.87)
- Image Watermarks are Removable using Controllable Regeneration from Clean Noise (ICLR 2025 · similarity 0.87)
- Hidden in the Noise: Two-Stage Robust Watermarking for Images (ICLR 2025 · 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.
