Presenters

Source

💧 Invisible Watermarking: How Meta is Protecting Creators at Scale 🚀

The internet is a vibrant ecosystem of content creators, and the ease of sharing and reposting videos has fueled incredible creativity. But it’s also created a challenge: how do we ensure creators get the credit they deserve when their work is shared and reused across the web? Meta is tackling this problem head-on with a fascinating solution: invisible watermarking (IW). Let’s dive into how they’re doing it, the hurdles they faced, and what the future holds!

The Problem: Content Recycling is Rampant 🌐

Think about your Instagram feed. How often do you see videos that have clearly been reposted from other platforms? It’s a common occurrence. Traditional methods of attribution – like visible watermarks or relying on metadata – just aren’t cutting it. Watermarks can be easily removed, and metadata often gets lost during video transcoding. This leaves creators vulnerable to having their work exploited without proper recognition or compensation.

Meta’s Solution: A Hidden Layer of Protection 🛡️

Meta’s answer? Invisible watermarking (IW). Imagine a digital fingerprint embedded within a video that’s undetectable to the human eye but can be reliably identified by software. This allows for robust content tagging even after a video has been edited or reposted. It’s a game-changer for creator attribution!

The Tech Journey: From GPUs to Optimized CPUs 👨‍💻

The team’s journey wasn’t a straight shot to success. Here’s a glimpse into the technical challenges they encountered:

  • GPU Experimentation - A False Start: Initially, the team explored using GPUs and neural networks to generate masks for the watermarking process. However, they quickly hit a wall. Parallelizing inference on GPUs resulted in dramatically increased latency, and the initial setup cost of loading the models onto GPUs was prohibitively time-consuming.
  • The CPU Pivot - And How They Won: After rigorous profiling and experimentation, the team made a strategic pivot to a CPU-based solution. While initially 2x slower than GPUs, they were able to achieve significant performance gains through careful optimization. Adjusting thread counts and sampling parameters brought performance to an acceptable level, proving that a CPU-based solution, with the right approach, could be a viable and operationally efficient alternative.
  • Scalability Testing – Finding the Sweet Spot: To ensure the system could handle the massive scale of Meta’s platforms, a robust load testing system was implemented. This helped the team determine the peak performance of the CPU-based solution before latency began to creep up.
  • Bit Rate Blues – Balancing Accuracy and Efficiency: Embedding the invisible watermark does increase the bit rate needed for video encoding. Initially, this increase was around 20%! The team ingeniously implemented a temporal sampling method to reduce this impact to a much more manageable 6%, striking a crucial balance between bit rate and detection accuracy.
  • Beyond Standard Metrics - A New Approach to Visual Quality: Standard visual quality metrics (like VMAF and SSIM) weren’t sufficient to accurately assess the quality degradation caused by the invisible watermark. The team relied on manual inspections and custom post-processing techniques to ensure the resulting videos looked great.

Key Takeaways & What’s Next ✨

This whole process has yielded some valuable lessons:

  • Don’t Dismiss the CPU: Sometimes, the most obvious solution isn’t always the best. Careful optimization can unlock surprising performance from CPUs.
  • Existing Metrics Need Improvement: The current suite of visual quality metrics isn’t always equipped to handle the nuances of technologies like invisible watermarking.
  • The Future is Bright (and Watermarked!): Meta’s ambitious goals for the future include:
    • Achieving 99.9% precision and >90% recall in invisible watermarking detection.
    • Developing more accurate and tailored visual quality metrics specifically for invisible watermarking.
    • Expanding the use of invisible watermarking to various video use cases, including tagging AI-generated content and further protecting the rights of original creators.

Invisible watermarking is a powerful tool for safeguarding creators and fostering a more equitable digital landscape. It’s a testament to the innovative thinking and dedication of the team at Meta, and a glimpse into the future of content protection online! 📡

Appendix