Robust AIGC Provenance with Invisible Watermarking

The explosive growth of AI-Generated Content (AIGC) has introduced unprecedented challenges in content authenticity, copyright attribution, and abuse traceability. As a responsible platform, TikTok makes every effort to control and govern what it has created for its users through GenAI features. We started in 2024 by joining C2PA and labeling AIGC with publicly verifiable manifests. Now we take a step further and introduce robustness to our AIGC provenance strategy. In this talk, we present TikTok’s approach to building a high-performance invisible watermarking system designed for production-scale AIGC services. Unlike metadata-based approaches (e.g., C2PA), which are trivially removable by file re-encoding or simple editing, invisible watermarks embed information directly into pixel-level content, offering far greater robustness against common attacks such as cropping, compression, resolution changes, and social media broadcast. Besides, it can be used as a backup retrieval mechanism for lost C2PA manifests. As of today, millions of AIGC created by TikTok are watermarked daily. TikTok manages to identify AI-generated content that was created by TikTok and re-uploaded by users by detection TikTok watermarks. This talk is intended for developers, product engineers, and technical leaders working on AI systems, content platforms who need to understand the practical implications of emerging AIGC regulations and how invisible watermarking can serve as a robust compliance solution.