Visual Fidelity vs. Robustness: Trade-Off Analysis of Image Adversarial Watermark Mitigated by SSIM Loss

Published: 06 Mar 2025, Last Modified: 16 Apr 2025WMARK@ICLR2025EveryoneRevisionsBibTeXCC BY 4.0
Track: tiny / short paper (3-5 pages)
Keywords: Adversarial watermark, AI watermark, Perturbation imperceptibility
TL;DR: Ultimately, adversarial watermark require both perturbations that are acceptable to users and effective defense. We aim to strike a balance between image fidelity and robustness against AI training in adversarial watermarks by leveraging SSIM Loss.
Abstract:

Adversarial watermark is an important technique for protecting digital images from unauthorized use and illegal AI training. However, conventional methods often introduce visually unpleasant artifacts, making the watermark easily perceptible. This results in an inherent trade-off between robustness and visual fidelity, where stronger protection comes at the cost of degraded image quality. In this work, we address this challenge by integrating SSIM loss into the perturbation embedding process using the Fully-trained Surrogate Model Guidance (FSMG) from baseline. By employing tunable SSIM weights, our approach balances the adversarial loss—designed to hinder unauthorized model training—with a perceptual loss that preserves image fidelity. Experimental results on CelebA-HQ and VGGFace2 show that our method effectively enhances image quality while preserving robustness, as validated by quantitative metrics and user evaluations confirming its practical viability for content protection.

Presenter: Jiwoo Choi
Format: Maybe: the presenting author will attend in person, contingent on other factors that still need to be determined (e.g., visa, funding).
Funding: Yes, the presenting author of this submission falls under ICLR’s funding aims, and funding would significantly impact their ability to attend the workshop in person.
Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.
Submission Number: 50
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview