Enhancing Deepfake Detection with a Hybrid CNN-BiLSTM Approach

Rapid improvements in deepfake technology have created serious threats to the integrity of digital information. Deepfakes have the potential to be employed for intimidating or blackmailing specific public figures, as well as for deceiving the general public. Concurrently, the advancement in deepfake technology has led to notable progress in developing detection algorithms. Primarily, these algorithms concentrate on analyzing images rather than delving into the temporal progression within video content. In the information era of today, detecting deepfakes has become more important than ever. The paper posits a pioneering hybrid approach for detecting deepfake videos, combining the scrutiny of human eye blinking patterns, utilizing Xception image characteristics, and implementing a Bidirectional Long Short Term Memory network (Bi-LSTM). This methodology leverages Convolutional Neural Networks(CNN) to extract image attributes and Bi-LSTM networks to scrutinize sequences consecutively. Moreover, it utilizes an adaptive spatio-temporal attention mechanism to concentrate on pertinent frames in the video sequences. The dataset used for experimental purposes includes both authentic and artificially generated videos from the Celeb-DF dataset. The proposed framework attained an accuracy rate of 86% in classification with an area under curve of 0.90, showcasing its efficacy in the identification of deepfake videos.