A Unified Deep Learning Framework for Face Forgery Detection in Varying Image Qualities Using TNVS-Net

Abstract

Face forgery is the manipulation or synthetic generation of facial images or videos to falsely represent someone's identity, expressions, or actions. Deep Learning (DL) methods have been widely used for both creating and detecting such forgeries. Various DL methods have been developed for face forgery detection and perform well on high-quality images and videos. However, real-world images are often compressed by social media platforms degrading visual quality, blurring fine details and introducing noise and artifacts that hinder detection quality. To address this, Multi-Scale Dual-Branch Network (MDCF-Net) is used which combines multi-scale Red Green Blue (RGB) and frequency domain features to detect manipulated faces. These features are processed through a Spatial Pyramid Pooling (SPP) layer to produce a fixed-length representation followed by a softmax classifier for prediction. But, MDCF-Net is mainly designed for detecting forgeries in highly compressed facial images. However, it has not yet attained the best possible performance in detecting forgeries on datasets that are not compressed or have modest compression. Hence, in this paper, Nexus Relation Network (NRN) branch along with RGB domain branch and a frequency domain branch of MDCF-Net is proposed. This proposed model termed as Three-Node Variable-Scale Network (TNVS-Net) has been developed for detecting highly compressed, low compressed or uncompressed facial images. The developed model is tested with three different datasets and found that it  achieves higher performance than the existing  classical face forgery prediction models.