We will develop a general DeepFake imagery detection algorithm that uses meta-learning framework to fuse multimodal cues (physiological, signal-level, data-level, and audio-level) to improve detection efficacy. Spatial localization will help identify locations that are more relevant to the detector. Temporal localization will help identify segments of frames to detect videos that are partially tempered. The Source Agnostic DeepFake Detector (SAFE) will provide interpretable results of the inconsistencies in imagery.
