Project R-16113

Training-free and model-agnostic approaches for generalizable and explainable AI-generated video detection. (Research)

Generative AI, primarily used for the generation of text and images, has recently succeeded in generating convincing videos. This new technology poses multiple threats to society and on an individual level, including election manipulation and impersonation of people. Research on detecting AI-generated videos is limited, focusing primarily on videos of human faces (deepfakes). Existing research on the detection of more general AI-generated videos is all data-driven, training a neural network on large amounts of data. These approaches suffer from a lack of generalizability, learning artefacts specific to the generative model used to create the training data. The aim of this PhD is to discover training-free and model-agnostic techniques, as have already been explored in the image domain, to distinguish between real and AI-generated videos. To reach this goal, I will also analyse the different artefacts present in videos generated by current video generation models. Finally, I will focus on the explainability of my proposed detector by researching the best ways to convey its decisions to users. The results of my PhD can be beneficial for multiple applications and societal stakeholders, including fact-checking in traditional and social media, digital forensics by police departments, cases related to national security by defence ministries and legislative entities at different levels.

01 November 2025 - 31 October 2029