PECCAVI embeds watermarks in specific regions of an image known as Non-Melting Points (NMPs). These strategically selected areas are the most crucial and resistant to adversarial attacks, including visual paraphrasing. The watermarking process begins by generating several paraphrased versions of the original image. For each version, the most important areas are identified using an advanced method called the XRAI algorithm. The most significant regions are then selected, and the system focuses on these areas while disregarding less important ones. Once the critical regions are identified, bounding boxes are drawn around them. These boxes are aggregated from all paraphrased versions and undergo a filtering process to select the most distinct, non-overlapping areas. This ensures that the watermark is placed in areas that are sufficiently large to be effective yet not too small to lose their impact. The approach also ensures that there are no overlapping boxes that could interfere with the watermark’s integrity. By embedding the watermark within these Non-Melting Points, PECCAVI ensures that the watermark remains robust and resistant to adversarial manipulation while preserving the overall visual quality of the image.

Once the Non-Melting Points (NMPs) are identified, PECCAVI employs several watermarking strategies to enhance resilience against attacks. Four distinct methods were tested: (1) baseline watermarking with state-of-the-art techniques such as ZoDiac, Stable Signature, and WAM, (2) adjusting the strength of the watermark, (3) single-channel watermarking, and (4) multi-channel watermarking. Among these, multi-channel watermarking proved to be the most effective, as embedding watermarks across multiple channels significantly increased resistance to attacks like visual paraphrasing. Additionally, varying the watermark strength, determined by the distance between the rings, further increased the system’s robustness, with smaller distances resulting in stronger, more resilient watermarks.