Frequency Domain Filtered Residual Network for Deepfake Detection

As deepfake becomes more sophisticated, the demand for fake facial image detection is increasing. Although great progress has been made in detection, performance of most existing methods degrade significantly when these are applied to detect low-quality images disappearance key clues during compression process. In this work, we mine frequency domain and RGB information specifically improve compressed images. Our method consists two modules: (1) a preprocessing module (2) classification module. module, utilize Haar wavelet transform residual calculation obtain mid-high joint fuse map with input. obtained by concatenation fed convolutional neural network classification. Because combination domain, robustness model greatly improved. extensive experimental results demonstrate that our approach can not only achieve excellent detecting images, but also maintain high-quality