Year: 2025
Type: article
Abstract: The increasing complexity of deepfake technology poses a serious threat to digital media integrity and strong detection and explanation mechanisms are needed. This project presents a complete system that detects not only deepfake videos but also gives detailed, interpretable explanations to support user understanding and trust. The detection pipeline integrates an InceptionV3-based Convolutional Neural Network (CNN) for spatial feature extraction and a Gated Recurrent Unit (GRU) for modeling temporal sequences, enabling accurate classification of video content as real or fake. Facial feature segmentation is performed using the CelebAMask-HQ model to highlight specific facial regions, while Grad-CAM is applied to visualize areas of high influence in the model's decision-making. Another new addition to the system is the integration of Gemini AI, which generates natural language explanations based on technical metrics such as confidence scores, face consistency, and temporal anomalies. The complete system is deployed via a Streamlit interface, offering an intuitive platform for users to upload videos and receive both technical and AI-powered insights. This work advances the field of explainable deepfake detection by combining deep learning, visual interpretation, and natural language explanation in a unified, user-friendly framework. Key Words: Deepfake detection, Explainable AI , InceptionV3, GRU, CelebAMask-HQ, Grad-CAM, Gemini analysis, Streamlit interface.
DOI: 10.55041/ijsrem47775
Authors: Artificial Intelligence and Machine Learning, S. Sri Harsha Varma
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4410380330