Revealing the Unseen: An Insightful Survey on Explainable AI for Deepfake Face Detection

A deepfake face is a technique that uses deep learning to create extremely realistic forged faces, making it challenging for humans to distinguish between real and fake faces. Existing deepfake face detection approaches show promise but heavily rely on complex deep neural networks (DNNs) and are often treated as black boxes. To address this, eXplainable AI (XAI) approaches have gained attention to enhance model interpretability and transparency. This survey reviews the most advanced XAI models used for detecting face deepfakes in videos and images. Our review examines the literature adapting to these XAI algorithms: Feature Visualization, Feature Attribution Methods, Attention Mechanisms, model-agnostic methods, interpretability through prototypes, and hybrid approaches. We perform a comparative analysis of different XAI methods and highlight their strengths and weaknesses. We also look ahead and highlight exciting, unsolved challenges that need to be addressed. This survey serves as a primary reference for researchers to enhance their understanding of neural networks using XAI techniques in the domain of deepfake detection in facial images and videos.