Towards Generalized Detection of Face-Swap Deepfake Images

As the prevalence of face-swap deepfakes on the Internet continues to rapidly increase, it is imperative for social media platforms to utilise a robust detection algorithm to identify these fake images in order to minimise the risk of harm from malicious content. However, there is currently no readily available detector to facilitate this process. This is mainly because face-swap detectors often 1) fail to correctly classify fake images generated by architectures that they have not encountered during training and are 2) highly susceptible to image manipulations such as compression techniques. In this paper, we present a novel approach for a deepfake image detector. Our approach adopts the EfficientNet-B4 Convolutional Neural Network architecture with noisy-student pre-training which addresses these issues through maximising the diversity of the training dataset and augmenting the input during training to ensure that the detector performs well on manipulated images. Our detector was tested against two different datasets containing a range of manipulated face-swaps created by DeepFaceLab, FSGAN and Faceswap. Our detector achieves an accuracy of 92.7% whilst maintaining a false positive rate below 1%. The results demonstrate that our proposed approach is effective at addressing the problems of generalisation which hamper efforts of deepfake detection.