Semi-Supervised Deep Domain Adaptation for Deepfake Detection

Abstract

With the advent and popularity of generative models such as GANs, synthetic image generation and manipu-lation has become commonplace. This has promoted active research in the development of effective deepfake de-tection technology. While existing detection techniques have demonstrated promise, their performance suffers when tested on data generated using a different faking technology, on which the model has not been sufficiently trained. This challenge of detecting new types of deepfakes, without losing its prior knowledge about deepfakes (catastrophic for-getting), is of utmost importance in today' s world. In this paper, we propose a novel deep domain adaptation frame-work to address this important problem in deepfake detection research. Our framework can leverage a large amount of labeled data (fake / genuine) generated using a particu-lar faking technique (source domain) and a small amount of labeled data generated using a different faking technique (target domain) to induce a deep neural network with good generalization capability on both the source and the target domains. Further, deep neural networks are data-hungry and require a large amount of labeled training data, which may not always be available in the context of deepfake de-tection; our framework can also efficiently utilize unlabeled data in the target domain, which is more readily available than labeled data. We design a novel loss function and use the stochastic gradient descent (SGD) method to optimize the loss and train the deep network. Our extensive empiri-cal studies on the benchmark FaceForensics+ + dataset, using three types of deepfakes, corroborate the promise and potential of our framework against competing baselines.