Towards efficient and effective face forgery detection

Abstract

Rapid progress in Face Forgery algorithms is becoming an increasingly relevant threat in the modern-day. To address the issue, multiple state-of-the-art forgery detection algorithms have been proposed. However, these models made use of deep and complex Convolutional Neural Networks (CNN) as their backbone, making them unrealistic to use in a production environment due to the high requirement for computational resources. Luckily, research on network pruning techniques has made tremendous progress in recent years, making it possible to compress CNN models by 90%, while retaining the same model accuracy. In this project, we experimented with and evaluated multiple pruning strategies and techniques. We then applied our findings by pruning a baseline face forgery detection model and achieved an impressive 2x speedup in the model inference time and a 70% decrease in model size. When our model is run under CPU-only hardware, a 30x speedup was achieved.