Debjyoti Ghosh

Deepfake generation has advanced to produce convincing synthetic audio and video, creating serious challenges for ensuring the credibility of digital content. This work proposes a scalable multimodal detection framework that leverages CNN-based transfer learning in combination with transformer-based fusion to capture spatial, temporal, and cross-modal inconsistencies. Video analysis incorporates face tracking, temporal dynamics, and artifact localization, while audio analysis employs Mel-Spectrogram features and pretrained embeddings. Additionally, speech–lip synchronization is evaluated to identify inconsistencies between phonemes and visemes. Experiments show that the system effectively distinguishes authentic content from manipulated variants (fake audio, fake video, or both), even under variations such as compression noise and resolution shifts. With applications in digital forensics, social media monitoring, and biometric verification, the proposed approach offers a robust defense against the growing sophistication of deepfake technologies.