Towards Solving the DeepFake Problem : An Analysis on Improving DeepFake Detection using Dynamic Face Augmentation

The creation of altered and manipulated faces has become more common due to\nthe improvement of DeepFake generation methods. Simultaneously, we have seen\ndetection models' development for differentiating between a manipulated and\noriginal face from image or video content. In this paper, we focus on\nidentifying the limitations and shortcomings of existing deepfake detection\nframeworks. We identified some key problems surrounding deepfake detection\nthrough quantitative and qualitative analysis of existing methods and datasets.\nWe found that deepfake datasets are highly oversampled, causing models to\nbecome easily overfitted. The datasets are created using a small set of real\nfaces to generate multiple fake samples. When trained on these datasets, models\ntend to memorize the actors' faces and labels instead of learning fake\nfeatures. To mitigate this problem, we propose a simple data augmentation\nmethod termed Face-Cutout. Our method dynamically cuts out regions of an image\nusing the face landmark information. It helps the model selectively attend to\nonly the relevant regions of the input. Our evaluation experiments show that\nFace-Cutout can successfully improve the data variation and alleviate the\nproblem of overfitting. Our method achieves a reduction in LogLoss of 15.2% to\n35.3% on different datasets, compared to other occlusion-based techniques.\nMoreover, we also propose a general-purpose data pre-processing guideline to\ntrain and evaluate existing architectures allowing us to improve the\ngeneralizability of these models for deepfake detection.\n