Robust DeepFake Face Detection Leveraging Xception Model and Novel Snake Optimization Technique

Abstract

DeepFake technology has created an existential crisis around authenticity in digital media with the ability to create nearly imperceptible forgeries on a massive scale, such as impersonating public figures for nefarious reasons like misinformation campaigns, harassment, and fraud. In this thesis, a model Xception is combined with the Snake optimization technique to ensure efficient and accurate detection of ADOR in practice. The former is deep CNN architecture Xception which exploits depthwise separable convolutions to perform efficient feature extraction, and the latter is a novel snake optimization that borrows lessons from real-life predatory snakes to dynamically adapt parameters for better exploration of search space while avoiding local optima. The combined modality is systematically evaluated using multiple challenging DeepFake video datasets and shows significant improvement. A comparison of performance with other methods showed that a mean accuracy, precision, recall, and F1-score was 98.53% for the Snake-optimized Xception model while outperformed some state-of-the-art approaches and traditional Xception itself. This helps in reducing missing of misdetection and reduction of false positives, helping achieve a tool that is highly effective for digital media forensics. Such discoveries open the door for this method to unlock new levels of digital content integrity, necessary in media verification and legal evidence authentication, as well as assist individuals dealing with fake news or videos attempting identity theft online. This research highlights the strong efficacy of coupling the Xception model with Snake optimization for DeepFake detection; thus, establishes a new state-of-the-art and will inspire future studies and applications to protect genuineness in digital media.