Deepfake Detection: A Multimodal Survey

The rapid advancement of generative artificial intelligence has catalyzed the emergence of deepfake technologies capable of cross-modal data fusion, posing systemic threats to digital security. To address these challenges, the academic community has developed multidimensional detection frameworks that integrate three core components: spatiotemporal consistency verification, cross-modal feature alignment, and semantic correlation inference. By synergistically processing multimodal data streams—including video, audio, and text—these frameworks leverage the complementarity and contradictions inherent in cross-modal features to identify forgery artifacts, substantially enhancing detection efficacy for sophisticated synthetic content. This study systematically examines the algorithmic architectures underpinning multimodal detection technologies, with focused analysis on optimized feature fusion strategies, innovative dynamic temporal modeling approaches, and cutting-edge adversarial training mechanisms. It further explores their application potential in critical scenarios such as political communication authentication and judicial digital forensics. The research confirms the paradigm's unique advantages in countering complex forgery attacks, establishing scalable technical pathways for developing intelligent defense systems against advanced deepfake threats.