Exposing Face-Swap Images Based on Deep Learning and ELA Detection

Abstract

The new development in Artificial Intelligence, has significantly improved the quality and efficiency in generating fake face images; for example, the face manipulations by DeepFake is so realistic that it is difficult to distinguish the authenticity-either automatically or by humans. In order to enhance the efficiency of distinguishing facial images generated by AI from real facial ones, a novel model has been developed based on deep learning and ELA detection, which is related to entropy and information theory, such as cross-entropy loss function in the final Softmax layer, normalized mutual information in image preprocessing and some applications of encoder based on information theory. Due to the limitations of computing resources and production time, DeepFake algorithm can only generate limited resolutions, resulting in two different image compression ratios between the fake face area as the foreground and the original area as the background, which would leave distinctive artifacts. By using the error level analysis detection method, we can detect the presence or absence of different image compression ratios, and then use CNN to detect whether the image is fake. Experiments show that the training efficiency of CNN model can be significantly improved by using the ELA method. And the detection accuracy rate can reach more than 97% based on CNN architecture of this method. Compared to the state-of-the-art models, the proposed model has the advantages such as fewer layers, shorter training time and higher efficiency.

Analysis of the current state of deepfake techniques-creation and detection methods

“…The model trained on the DFDC dataset obtained 91.5% accuracy, a loss value of 0, and AUC of 0.91. In 2019, Zhang and Zhao [17] proposed a new deep learningbased method for identifying AI face photos from real-world facial images. Artificial intelligence (i.e., AI) facial recognition has been improved by using a new model based on deep learning and detection-level analysis.…”

“…[37] MarioNETte Identity adaption does not necessitate a further fine-tuning process. [17] DiscoFaceGAN Adopt 3D priors in adversarial.…”

A survey of deepfakes in terms of deep learning and multimedia forensics

“…So why do we not exploit the advantages of multimedia forensic methods and deep learning methods to get algorithms more robust in terms of detecting deepfakes?. For example, detecting different ratios of image compression using error level analysis followed by using the CNN model increases the detection accuracy [61]. This is because images with JPEG formats have the same compression level, so manipulation applied to that image will disturb the pressure levels between the modified area and the surrounding areas.…”

Media Forensics Considerations on DeepFake Detection with Hand-Crafted Features

“…Looking for neural network-based approaches implementing such a feature space, the papers of Zhang et al [ 52 , 53 ] have to be mentioned here. In contrast to our approach, they developed an automatic approach using a CNN.…”

“…If the CNN is able to extract these counterfeit features, then the input image of the CNN is a DeepFake. Even though the detection in [ 52 , 53 ] uses only DeepFake images in its tests, it would be possible to upgrade this approach for a DeepFake detection of videos.…”