Explainable Multimodal AI for Deepfake Detection and Digital Content Authenticity

Abstract: Recent breakthroughs in AI and deep learning have enabled the creation of highly convincing synthetic media — commonly termed deepfakes — that are increasingly difficult to distinguish from authentic content. Techniques such as GANs, transformer-based architectures, and diffusion models introduce serious risks to cybersecurity, journalistic integrity, digital trust, and democratic processes. Existing solutions are predominantly limited to single-modality analysis, hindering effectiveness against coordinated multimedia misinformation spanning video, audio, and text simultaneously. This paper surveys existing literature on deepfake detection, explainable AI, and multimodal misinformation analysis, identifying key research gaps and limitations. Based on these findings, we propose an Explainable Multimodal AI Framework that unifies ResNet18, CNNs, and DistilBERT with SHAP, LIME, and RAG-based contextual reasoning for simultaneous detection and authenticity verification of manipulated multimedia content. The proposed framework is expected to deliver improved accuracy, transparency, and real-time inference capability over existing unimodal and non- explainable approaches when evaluated on standard benchmarks such as FaceForensics++, DFDC, ASVspoof 2019, and FakeNewsNet.

Keywords: Deepfake Detection; Explainable AI; Multimodal Learning; Digital Authenticity; Generative Adversarial Networks; CNN; ResNet18; DistilBERT; Retrieval-Augmented Generation; SHAP; LIME; Real-Time Detection.