Publication

Abstract : The increasing complexity of deepfake technologies necessitates the development of detection methods to counter their harmful propagation. This research introduces a ground-breaking approach, to detecting deepfake images by utilizing Hierarchical Spatial Temporal Contextual Embedding Networks (HST CEN) in combination with Dynamic Attention Mechanisms (DAM).By addressing the challenges in understanding temporal context within deepfake images HST CEN incorporates multi level contextual embeddings that capture subtle spatial and temporal cues. At the time DAM employs evolving attention mechanisms that hierarchically identify important features at multiple levels enabling robust detection of manipulated imagery.The methodology includes an evaluation on the ForgeryNet dataset showcasing the superiority of the model in discerning deepfake images compared to architectures. The results demonstrate performance leading to improved accuracy, specificity and generalization across various forgery techniques and different identity contexts. The qualitative analysis reveals attention value distributions providing insights into how the model identifies manipulated features for nuanced analysis, by end users. The impacts of this detection method have implications, in scenarios providing a strong defense against the spread of sophisticated deepfake content. The suggested HST CEN with DAM not enhances the capabilities of identifying manipulated images but also highlights the importance of having a deep understanding of spatial temporal context and utilizing dynamic attention mechanisms to combat the widespread use of artificially created visuals, in todays digital world.