Exploiting Deepfakes by Analyzing Temporal Feature Inconsistency

Abstract

In recent years, the rapid advancement of image generation technology has facilitated the creation of counterfeit images and videos, posing significant challenges for content authenticity verification. Malefactors can easily extract videos from social networks and generate their own deceptive renditions using state-of-the-art techniques. The latest Deepfake face forgery videos have reached an unprecedented level of sophistication, making it exceptionally difficult to discern signs of manipulation. While several methods have been proposed for detecting fraudulent media, they often target specific aspects, and as new attack methods emerge, these approaches tend to become obsolete. This paper presents a novel detection approach that combines Convolutional Neural Networks (CNN) and Long ShortTerm Memory Networks (LSTM). Initially, CNN is employed to extract image features from each frame of the input facial video, capturing subtle alterations and irregularities in manipulated content. Subsequently, the extracted feature sequence is used to train the LSTM network, mimicking the temporal consistency of human visual perception and enhancing the effectiveness of counterfeit video detection. To validate this methodology, a comprehensive evaluation is conducted using the FaceForensic++ dataset, affirming its proficiency in identifying Deepfake counterfeit videos.