A Hybrid Explainable Artificial Intelligence Framework for Deepfake Attribution and Cross-Platform Disinformation Campaign Tracking

Abstract

Abstract The rapid advancement of Generative Artificial Intelligence (GenAI) has accelerated the creation and dissemination of highly realistic deepfakes, posing significant cybersecurity threats through identity fraud, misinformation, political manipulation, social engineering, and coordinated cross-platform disinformation campaigns. Although existing deep learning-based deepfake detection systems achieve high classification accuracy, most are limited to binary detection (real versus fake) and provide little support for source attribution, explainability, cross-platform campaign analysis, or digital forensics. These limitations reduce their effectiveness in cyber threat intelligence and forensic investigations. This study aimed to develop a Hybrid Explainable Artificial Intelligence (XAI) Framework for Deepfake Attribution and Cross-Platform Disinformation Campaign Tracking. The objectives were to develop a multimodal attribution framework for images, videos, audio, text, and metadata; integrate Explainable Artificial Intelligence techniques for transparent decision-making; implement graph-based cross-platform disinformation tracking; incorporate cyber threat intelligence and digital provenance; and evaluate the framework against existing approaches. A Design Science Research (DSR) methodology was adopted. The framework was implemented using Vision Transformers (ViT), Temporal Transformers, Wav2Vec 2.0, RoBERTa, Random Forest, XGBoost, and Multilayer Perceptron (MLP) within a weighted ensemble architecture. SHAP, LIME, and Grad-CAM were employed for explainability, while Graph Attention Networks (GATs) and Temporal Graph Neural Networks (TGNNs) enabled campaign tracking. Evaluation was conducted using FaceForensics++, Celeb-DF, DFDC, FakeAVCeleb, ASVspoof, Fakeddit, LIAR, CoAID, Twitter/X, and PHEME datasets. The proposed framework achieved 92.1% accuracy, 91.2% precision, 90.9% recall, 91.0% F1-score, and 95.9% ROC-AUC. It further attained 90.8% attribution accuracy, 98.3% Top-5 attribution precision, 94.2% explanation fidelity, 93.4% community detection accuracy, and an end-to-end inference latency of 63.8 ms, demonstrating near real-time performance. The study concludes that integrating multimodal attribution, Explainable AI, graph analytics, cyber threat intelligence, and digital forensics significantly enhances deepfake attribution and coordinated disinformation tracking beyond conventional detection systems. The proposed framework provides an effective solution for cybersecurity operations, digital forensics, law enforcement, social media monitoring, and national security applications. Keywords: Deepfake Attribution, Explainable Artificial Intelligence, Cross-Platform Disinformation, Cyber Threat Intelligence, Digital Forensics, Graph Neural Networks, Multimodal Learning, Cybersecurity