Detecting Compressed Deepfake Images Using Two-Branch Convolutional Networks with Similarity and Classifier

As a popular technique for swapping faces with someone else’s in images or videos through deep neural networks, deepfake causes serious threat to the security of multimedia content today. However, because counterfeit are usually compressed when propagating over Internet, and compression factor used is unknown, most existing detection models have poor robustness unknown factors. To solve this problem, we notice that an image has high similarity its based on symmetry, not easily affected by factor, so feature can be as important clue detection. A TCNSC (Two-branch Convolutional Networks Similarity Classifier) method combines independence proposed paper. The learns two representations from image, i.e., counterpart authenticity image. joint training strategy then utilized extraction, which characteristics obtained learning while obtaining characteristics, model trained robust learning. Experimental results FaceForensics++ (FF++) dataset show significantly outperforms all competing methods under three settings high-quality (HQ), medium-quality (MQ), low-quality (LQ). For LQ, MQ, HQ settings, achieves 91.8%, 93.4%, 95.3% accuracy, state-of-art (Xception-RAW) 16.9%, 10.1%, 4.1%, respectively.