Faceswap Finder: A Fusion-Based Deepfake Detection Technique

The significant improvement in deepfake generation algorithms has made possible the manipulation of visual data. The advancement of GAN technology and tools has made it relatively simple to use source and target photos and build realistic deepfakes. The challenges associated with deepfakes include political defamation, impersonation, misinformation, and cyberstalking. The generation of deepfakes through faceswap approach hinders the detection oftentimes as it spans numerous artifacts – illumination, facial skin tones, and other conditions. Existing techniques do not employ robust mechanisms to detect such deepfakes either due to computational inefficiency or focusing on a few parameters for decision-making. To counteract these dangers, a robust deepfake detection model must be developed and implemented. This paper proposes Efficient-Fused Net (EF-Net) - a transfer learning approach to two EfficientNet-B5 models with partial freezing between its layers that can differentiate between real and fake images generated from faceswap implementation. We evaluated our model’s performance using the challenging deepfake subset of the extensive and varied FaceForensics++ and DFDC-Preview datasets. The proposed method outperforms state of the art methods in the detection of faceswap deepfakes, surpassing the performance by $89.19 \%$ accuracy on DFDC-Preview and $\mathbf{9 7 . 7 0 \%}$ of Deepfakes subset of FaceForensics++.