AGATE: Stealthy Black-Box Watermarking for Copyright Protection of Multimodal Models

AGATE is a model-agnostic black-box backdoor watermarking framework for copyright protection of multimodal models, addressing the challenges of watermark stealthiness and robustness—existing methods are easily maliciously detected and forged by adversaries, leading to watermark escape. Methodologically, the paper proposes an adversarial trigger generation method that generates stealthy adversarial triggers from ordinary datasets, inducing semantic shift while maintaining visual fidelity; to mitigate anomaly detection in model outputs, it proposes a post-transform module that corrects model outputs by narrowing the distance between adversarial trigger image embeddings and text embeddings; it then proposes two-stage watermark verification, comparing the two sets of results with and without the transform module to determine whether the current model infringes copyright. Experiments show that AGATE consistently outperforms state-of-the-art methods on five datasets across two downstream tasks, multimodal image-text retrieval and image classification, and robustness is verified under two adversarial attack scenarios.