A Hybrid MobileNet-LSTM Model for Enhanced Detection of Deepfake Media in Real-Time Image and Video Analysis
Neereddula Subhasri, Midde Gangadhar, Motte Divakar, Neelam Ediga Bhanu Chandra Goud, M. Ramaraju, Kilari Sreenivasulu · 2025
Deepfakes, or Al-generated manipulated media, pose a significant threat to media integrity, privacy, and public trust. Detecting deepfakes effectively remains a challenge due to their increasingly sophisticated nature. This research addresses the issue of detecting both spatial and temporal inconsistencies in deepfake images and videos by proposing a hybrid MobileNet-LSTM model. The MobileNet component extracts spatial features such as texture anomalies and facial misalignments, while the LSTM component analyzes temporal dependencies to identify inconsistencies in video sequences, such as unnatural facial expressions or motion patterns. The hybrid model is trained on a large and diverse dataset of real and deepfake media, ensuring its adaptability to emerging deepfake techniques. The outcome of this research is a highly accurate, real-time deepfake detection system that performs with an accuracy of 91.8%, achieving improved precision (89.4%), recall (90.5%), and F1-score (89.9%) compared to existing baseline models. The model’s lightweight design ensures real-time performance, making it suitable for deployment in social media monitoring, digital forensics, and law enforcement. This work contributes to the ongoing efforts in combating misinformation and enhancing digital media authenticity by providing a scalable and efficient deepfake detection solution.