Reliability-Aware Multi-Representation Fusion for Detecting Deepfake Images on Social Media Platforms

Abstract

As generative AI continues its rapid adoption, the creation of synthetic face images is becoming increasingly indistinguishable from real content on social media platforms, leading to the potential of misinformation, loss of trust, and social disruption. High accuracy of deep learning models is gained, but they tend to lose accuracy when the image quality or manipulation type or generation source changes from platform to platform. While multi-representation approaches are highly complementary, using RGB and error-level analysis, noise and frequency cues, they are predominantly based on fixed fusion weights that do not take into account the reliability of each cue with respect to a specific image. This study presents the Reliability-Aware Multi-Representation Fusion Network (RAM-FusionNet) that uses dedicated Xception branches to extract features and trainable, input-dependent reliability-gating mechanism to adaptively weight each representation. RAM-FusionNet achieves accuracy of 98.1% (up from 95.8% for RGB baseline), AUC of 0.996 (up from 0.984 for RGB baseline), good generalization on unseen datasets, good robustness to common social-media perturbations, and good calibration. The reliability-aware fusion is 0.4-point superior over fixed fusion, and also retains the cross-dataset accuracy rate of 91.7% under medium JPEG. The findings validate the effectiveness of adaptive and reliability-driven fusion in bolstering trustworthy deepfake detection, paving the way for more robust digital information systems and decisions regarding technology-fueled social change.