1/15成本，实现AI水印新SOTA | 南洋理工大学&A*STAR

Core Viewpoint - The article discusses the emerging consensus in the industry regarding the need for watermarking AI-generated content to ensure traceability, highlighting the limitations of traditional watermarking methods and introducing a new approach called MaskMark developed by researchers from Nanyang Technological University and A*STAR [1][3]. Group 1: Limitations of Traditional Watermarking - Traditional watermarking methods treat images as a whole, leading to failures in watermark extraction when parts of the image are altered, and they cannot protect specific areas like faces or logos [2]. - The inability to protect specific regions poses a significant challenge for content verification and copyright protection [2]. Group 2: Introduction of MaskMark - MaskMark is a novel local robust image watermarking method that significantly outperforms the state-of-the-art model WAM from Meta, with a training cost only 1/15 of WAM [4][5]. - The core idea of MaskMark is to inform the model where the watermark is embedded, allowing for precise insertion and extraction [5]. Group 3: Technical Features of MaskMark - MaskMark has two versions: MaskMark-D (decoding mask) and MaskMark-ED (encoding and decoding mask), focusing on dual optimization during training and inference [6]. - It supports multiple watermark embeddings, precise localization of tampered areas, and flexible extraction of local watermarks, adaptable to various bit lengths (32/64/128 bits) [7][8]. Group 4: Performance Metrics - MaskMark demonstrates high extraction accuracy, maintaining nearly 100% bit accuracy even under high visual fidelity conditions (PSNR > 39.5, SSIM > 0.98) [13]. - In local watermarking tasks, MaskMark outperforms existing global methods and the leading local watermark model WAM, especially in small area embedding scenarios [14][18]. Group 5: Efficiency and Scalability - MaskMark's training process is efficient, requiring only about 20 hours on a single A6000 GPU, with a computational efficiency (TFLOPs) 15 times higher than WAM [22]. - The method allows for easy scalability to different bit lengths while maintaining high performance, unlike WAM which is limited to 32 bits [20].