Exploiting spatiotemporal inconsistencies to detect deepfake videos in the wild
Atharva Khedkar, Atharva Peshkar, Ashlesha S. Nagdive, Mahendra A. Gaikwad, Sudeep Baudha · 2022 10th International Conference on Emerging Trends in Engineering and Technology - Signal and Information Processing (ICETET-SIP-22) · 2022
Cyberspace is an emerging battlefield and deepfakes are being constantly weaponized by malicious actors. With rapid advancements in media synthesis technologies, detecting deepfakes is becoming increasingly difficult. The following paper presents a unified approach focusing on the fusion of Convolutional Neural Networks and Long Short Term Memory Networks for spatial and temporal analysis of deepfake videos. This study compares the performance of the most prevalent and frequently used deepfake detection methods- convolutional neural networks (CNN) and convolutional neural networks combined with long-short term memory networks (CNN-LSTM) with our architecture on a combined dataset consisting of videos from Face Forensics++ and Deepfake Detection Challenge Dataset, which consists of multiple types of manipulated media- Deepfakes, Faceswaps, Neural Textures, Face Shifter and Face2Face. We find that our architecture provides a 2.5% increase in detection accuracy over the most frequently used current deepfake detection method (CNN-LSTM).