Hybrid Deep Learning Technique-Based Deepfake Detection and Multimedia Forensics

The proliferation of AI-generated synthetic media, commonly known as deepfakes, poses significant threats to digital trust, political integrity, and multimedia forensics. Existing detection methods often suffer from limited generalizability across manipulation types, vulnerability to compression artifacts, and inadequate exploitation of both spatial and temporal cues. In this paper, we propose the Hybrid Deep Learning for Multimedia Forensics (HDL-MF) framework, a novel dual-stream architecture that jointly leverages a convolutional spatial encoder for fine-grained forgery artifact extraction and a bidirectional gated recurrent unit (Bi-GRU) temporal stream for modeling inter-frame inconsistencies. A cross-modal attention fusion module integrates spatial and temporal representations at multiple granularities, while a temporal contrastive learning (TCL) objective enforces discriminative embedding separation between authentic and manipulated video sequences. We further incorporate a nine-channel forensic feature map—comprising local binary patterns, Sobel and Laplacian gradients, CIE LAB color residuals, HSV illumination inconsistencies, Gabor skin-texture responses, and JPEG blocking artifact maps—as an explicit supervisory signal. Extensive experiments on FaceForensics++ (FF++), Celeb-DF v2, and OpenForensics demonstrate state-of-the-art performance, achieving 98.6% accuracy and 99.1% AUC on FF++ and 95.2% accuracy on Celeb-DF v2 under cross-dataset evaluation. Ablation studies confirm the contribution of each proposed component.