AI Paper Atlas

Research map / Adversarial robustness and security / Certified and empirical robustness

Certifiably Robust Image Watermark

Zhengyuan Jiang, Moyang Guo, Yuepeng Hu, Jinyuan Jia, Neil Zhenqiang Gong

ECCV 2024 · Adversarial robustness and security › Certified and empirical robustness · approach: Randomized smoothing certification · 13 citations (OpenAlex)

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 changedWhereMeasuredEffect
use regression smoothing instead of multi-class or multi-label smoothing (ablation)smoothed watermarking · Stable Diffusion, certified robustnessCFNR and CFPRImproved
train the base watermark encoder/decoder with Gaussian-noise adversarial training instead of standard training (ablation)regression smoothing watermark · Stable Diffusion, certified robustnessCFNR and CFPR: SSIM 0.943 vs 0.941 (similar quality)Improved
lower the detection threshold tau (ablation)regression smoothing watermark · Stable Diffusion, certified robustnessCFNR and CFPRMixed
set smoothing noise sigma too small or too large (ablation)regression smoothing watermark · Stable Diffusion, certified robustnessCFNR and CFPRHurt
apply randomized smoothing to the watermark decoderHiDDeN base watermark · JPEG compression removal attackFNR: 0.25 -> close to 0Improved
apply randomized smoothing to the watermark decoderHiDDeN base watermark · black-box removal/forgery attackperturbation size at given query budgetImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy