A LIGHTWEIGHT EXPLAINABLE DEEPFAKE DETECTION FRAMEWORK USING SELF-BLENDED AUGMENTATION AND MULTI-DOMAIN FEATURE FUSION

Abstract

Deepfake detectors trained only on manipulation-specific artefacts often lose accuracy when tested on unseen datasets and compression settings. This paper proposes a lightweight and explainable framework for facial video deepfake detection that combines self-blended augmentation, spatial artefact learning, frequency-domain analysis, and physiological consistency cues. The proposed design uses self-blended images to synthesise hard negative samples from pristine faces, encouraging the detector to learn generic forgery boundaries rather than memorising dataset-specific manipulation traces. A compact convolutional backbone extracts spatial features, while frequency residual maps and remote photoplethysmography-inspired temporal descriptors provide complementary evidence for manipulated content. A gated multi-domain fusion block integrates the feature streams and an explainability module produces Grad-CAM heatmaps and domain-level contribution scores. The framework is designed for resource-constrained deployment while retaining interpretability for forensic analysts and platform moderators. A full experimental protocol is proposed using FaceForensics++, Celeb-DF, DFDC, and cross-dataset validation. The expected contribution is a practical architecture that improves generalisation, reduces computational overhead, and provides transparent evidence for its predictions.