Implementation-Free Forensic Watermarking for Adaptive Streaming with A/B Watermarking

Abstract

Forensic watermarking enables the identification of digital pirates after they illegally re-distribute copyright-protected videos. For adaptive streaming, these methods are best combined with A/B watermarking, in which two watermarked versions are created for each video segment, and subsequently mixed in order to create a large number of uniquely watermarked videos. Although good video quality and low bitrate are key characteristics of a good watermarking system, existing methods objectively lower the compression efficiency. Additionally, they often require complex implementations. Therefore, this paper proposes an implementation-free, rate-distortion-preserving watermarking technique to be used with the scalable A/B watermarking concept. Even though the embedding is performed during compression, it does not change the existing video encoder implementation. Instead, it only changes the target bitrate parameter in order to create different compression artifacts. These artifacts represent the watermark but are not noticeable due to high-quality coding. As such, the rate-distortion performance is nearly equal to that of ordinary, unwatermarked compression (i.e., a BD-rate of 0.02% and $-0.10$% when applied with H.264/AVC and H.265/HEVC encoders, respectively). Furthermore, the robustness is equal or better than state-of-the-art methods with comparable embedding complexities. More specifically, in case of recompression attacks, nonzero false negative rates are only reported when a watermarked video is initially compressed with a high quality and degraded to a very low quality. Consequently, the proposed scheme can be used in practice by adaptive-streaming platforms without a quality decrease, bitrate increase, or implementation overhead.