Many perceptual hashing algorithms for videos have been proposed and implemented. This white paper defines a comprehensive benchmark for analyzing the effectiveness of perceptual hash algorithms and systems to videos, and presents the results for several systems in this context.
The benchmark evaluation of video hash algorithms is designed around several goals:
- Improved results for all;
- Feedback-based design;
- Testing of diverse scenarios;
- Repeatable by third parties.
The Tech Coalition is committed to advancing tools and technologies that enhance online safety, particularly in combating child sexual abuse material (CSAM). Funding research like this allows us to develop benchmarks that guide industry in improving the accuracy, efficiency, and reliability of perceptual hashing algorithms.
We extend our gratitude to Brian Levine and his team at the University of Massachusetts Amherst’s Rescue Lab, Fausto Morales, Meta, Microsoft, Thorn, Videntifier, and all contributing members for their support in this project.