TÜBİTAK 3501 Project

The objective of this project is to develop a multimodal and modular deepfake detection platform capable of holistically identifying various types of deepfake manipulations. Most existing studies focus either on scenarios where the entire video is fake or target a single type of manipulation. However, in real-world scenarios, partial manipulations—such as adding fake content to specific segments of an authentic video (splicing) or removing an existing object (inpainting)—are more prevalent and significantly more challenging to detect.

The primary scientific contribution of the project is built upon a hierarchical decision mechanism that integrates three distinct detection approaches into a single platform: (1) temporal inconsistency analysis, (2) image-level splicing/inpainting detection, and (3) audio-visual desynchronization. The project is designed to operate across three different levels of difficulty: (i) detection of fully fake videos, (ii) detection of fake content added to an authentic video, and (iii) as the most challenging scenario, frame-level localization of short fake segments interspersed within the frames of a normal video. In this context, a unique and challenging dataset will be created due to the inadequacy of existing datasets.