Generalizable deepfake video detection for multimedia security via temporal–spatial artifact localization and adaptive sequence modelling

Due to the rise in face manipulation algorithms, it is difficult to distinguish real videos from computer-generated videos.For this, the paper proposes HABSNet, a model that performs artifact localization and temporal consistency analysis.The facial regions identified by MTCNN are analysed by ResNeXt-50 backbone model with the help of two-stream attention mechanism for detecting blending and frequency artifacts, as well as a BiLSTM based inter-frame self-attention gate for detecting unnatural eye movements and head pose changes.Trained on the FaceForensics++ (C23) dataset, the HABSNet achieves an accuracy of 94.88% and generalizes well on Celeb-DF v2 (0.950) and DFDC (0.920).For multimedia security the proposed framework contributes by allowing identification of manipulated videos across varied deepfake generation techniques.