DEEPSHIELD AI: DEEPFAKE DETECTION USING VISUAL INCONSISTENCIES

Highly Accelerated breakthroughs in the area of deep learning media generation have given rise to the production of highly realistic deepfakes, thus threatening the authenticity of information, personal privacy, and trust in the digital realm.Although original approaches to deepfake detection were based solely on visual cues, recent breakthroughs in generation models have made such approaches less efficient.This manuscript is based on research in the area of audio-visual inconsistencies, and the author has developed the DeepShield AI videobased deepfake detection platform based on modeling the audio-visual inconsistencies in facial motion.According to the proposed platform, a pretrained convolutional network can be utilized for spatial feature extraction on the facial area, and a long short-term memory network can be utilized for modeling the temporally corresponding facial motions.Performance analysis on various datasets shows the efficiency of the proposed platform, based on modeling facial motions, for the discrimination of original and manipulated facial videos.Although the current platform is based on visuals only, it has the potential to establish a robust, state-ofthe-art platform for the detection of deepfakes with the addition of audio-visual fusion analysis.