Bayesian Inference for Robust Video Watermarking
Track: tiny / short paper (3-5 pages)
Keywords: video watermarking, bayesian inference, video compression
TL;DR: We propose a plug-and-play Bayesian extractor that recovers video watermarks under compression without modifying existing embedding models.
Abstract: We propose a simple yet effective Bayesian extractor for multi-frame video watermarking that can be plugged into any existing image-based watermarking method, such as HiDDeN, CIN, MBRS, TrustMark, WAM, or VideoSeal.
In particular, we focus on challenging real-world conditions where videos undergo repeated or strong compression (e.g., H.264, H.265) or frame-rate changes that typically degrade watermark signals severely.
When all frames carry the same hidden bits, our Bayesian extractor treats each frame’s output as an independent observation and aggregates the log-likelihood ratios across frames, in contrast to naive averaging.
Despite only modifying the extraction phase, this approach consistently boosts bit accuracy under moderate-to-aggressive compression, frame-rate conversions, and other distortions—while preserving the same watermark imperceptibility and embedding efficiency as the baseline.
Experiments on diverse transformations and watermarking models show that these benefits are particularly pronounced when frames encounter uneven or heavy distortions, making our Bayesian extraction a lightweight but potent upgrade for robust video watermarking.
Presenter: ~Wonhyuk_Ahn1
Format: Yes, the presenting author will attend in person if this work is accepted to the workshop.
Funding: Yes, the presenting author of this submission falls under ICLR’s funding aims, and funding would significantly impact their ability to attend the workshop in person.
Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.
Submission Number: 18
Loading