A Comparative Study of Xception-LSTM and Xception-GRU for Deepfake Detection in Journalism

This research addresses the growing threat of deepfakes, AI-generated synthetic media, to journalism and public trust in digital content. Accurately detecting manipulated videos and images remains a critical challenge, especially under computational and data constraints. The study focuses on data-efficient deepfake detection by comparing two hybrid deep learning architectures: Xception combined with Long Short-Term Memory (LSTM) and Xception with Gated Recurrent Units (GRU), using the Deepfake Detection Challenge (DFDC) dataset. The paper evaluates model performance in terms of accuracy, computational efficiency, and data requirements, highlighting trade-offs between model complexity and practical usability. Findings indicate that GRU-based models are more computationally efficient and generalize better with limited GPU resources, making them suitable for real-world journalistic applications. This research contributes valuable insights into developing effective and practical deepfake detection systems, aiming to improve media integrity and counter misinformation.