CoMSMark watermarking blocks multi-screen collusion with 90% accuracy

Screen-shooting remains a critical security risk for confidential information. Existing watermarking solutions protect copyright, but when the same watermarked image appears on multiple screens, the shared watermark patterns become predictable and vulnerable to a multi-screen collusion attack—where attackers combine footage to remove or forge the watermark. To address this, researchers Mingyue Chen, Xin Liao, Yufeng Wu, Han Fang, and Xiaoshuai Wu have developed CoMSMark, a framework that conditions watermark generation on each screen's unique identifier. CoMSMark uses a style modulation mechanism that lets the encoder assign screen-specific residual patterns to the watermark, making source attribution reliable and preventing attackers from deriving a common signal across screens. Additionally, a collusion suppression loss is introduced to minimize shared residual components across different screens, while encouraging the model to output high-entropy predictions for forged samples, effectively neutralizing collusion-based forgery attempts.

The framework operates with an image-agnostic encoding paradigm, meaning it generates watermark residuals independently of the image content. This allows for efficient large-scale distribution without need per-image optimization. Extensive testing shows that CoMSMark delivers strong anti-collusion performance: it maintains an average watermark accuracy above 90% under removal attacks, while forged-watermark accuracy stays near 50%—essentially random, proving that forged samples are not accepted. The system also exhibits competitive robustness under practical screen-shooting scenarios, including varied capture distances and angles. CoMSMark represents a significant step forward in digital watermarking, specifically tailored to the multi-screen environments common in modern offices, conference rooms, and public displays. By eliminating the shared residual patterns that enable collusion, it sets a new standard for protecting confidential visual data.