FakeTalkerDetect: Effective and Practical Realistic Neural Talking Head Detection with a Highly Unbalanced Dataset

Detecting realistic fake images and videos is an increasingly important and urgent problem because they can be maliciously used. In this work, we propose FakeTalkerDetect, which is based on siamese networks to detect the recently proposed realistic talking head with few-shot learning. Unlike conventional methods, we propose to use pre-trained models with only a few real images for fine-tuning in siamese networks to effectively detect the fake images in a highly unbalanced data setting. Our FakeTalkerDetect achieves the overall accuracy 98.81% accuracy in detecting fake images generated from the latest neural talking head models. In particular, our preliminary work also demonstrates the effectiveness for the highly unbalanced dataset.