TNVS-Net with Context-Aware Relation Network for Robust Face Forgery Detection across Varying Image Qualities

Abstract

Face forgery is the manipulation or synthetic generation of facial images or videos to falsely represent a person’s identity, expressions, or actions. Deep Learning (DL) techniques have been extensively utilized for both generating and detecting such forged media. Several DL-based face forgery detection models have achieved significant performance on high-quality facial images and videos. However, in real-world scenarios, facial images are often compressed by social media platforms, which reduce image quality, obscure fine visual details, and introduces artifacts and noise that negatively affect detection accuracy. To overcome these challenges, a Multi-Scale Dual-Branch Network (MDCF-Net) was introduced, which integrates multi-scale Red Green Blue (RGB) features and frequency domain information for effective forgery detection. The extracted features are processed through a Spatial Pyramid Pooling (SPP) layer to generate fixed-length feature representations, followed by a softmax classifier for final prediction. Although MDCF-Net performs effectively on highly compressed facial images, its performance is limited when handling low-compressed or uncompressed datasets. Therefore, this paper proposes a Context-Aware Relation Network (CARN) branch integrated with the RGB domain branch and frequency domain branch of MDCF-Net. The proposed architecture, named Three-Node Variable-Scale Network (TNVS-Net), is designed to detect forged facial images across highly compressed, low-compressed, and uncompressed images. Experimental evaluation on three benchmark datasets demonstrates that the proposed TNVS-Net achieves superior performance compared to existing classical face forgery detection models.