Abstract:
"Deepfakes refer to altered digital content through deep generative models to achieve photorealistic standards. Deepfake instances related to celebrities are growing due to the accessibility of large quantities of data. Recently, several deepfake detection approaches have been developed to focus on identifying fake faces from genuine faces when they are tested on similar training forgery patterns. Somehow, these methods' performance is limited when tested on unseen patterns.
In this research, the author proposes a deepfake detection method to increase the generalization of the model by detecting both pixel-level and noise-level manipulations to make the final prediction. The proposed network has two methods for extracting features. First, there's a pixel-level classifier that detects pixel-level changes using convolutional neural networks. Second, there's a noise-level classifier that identifies noise-level changes based on the Peak Signal-to-Noise Ratio (PSNR) to verify the authenticity of celebrity images. After that, a logical method was developed to compare the results from both classifiers. If either classifier detects the image as 'fake,' the final output will be labelled as fake, covering both pixel and noise manipulations.
The proposed method extracts forgery traces caused by pixels and noise manipulations suggesting to have a good generalization ability. Extensive experimental results on three widely used datasets show that the proposed method achieves better generalization performance against unseen forgery patterns."
