Neutralizing Misleading Visual Cues: An Attention-Driven Encoder–Decoder Framework with Auxiliary Forensic Reconstruction for Video Deepfake Detection

As deepfake generation methods are constantly evolving, there is a need for new methods that can detect subtle manipulation artefacts in video content, both in terms of spatial and temporal characteristics. In this study, an encoder–decoder architecture is proposed, where the classification and forensic reconstruction are performed simultaneously at the encoder and decoder side, respectively. The architecture consists of a six-layer convolutional encoder to obtain spatial representations from uniformly sampled facial frames of 128 × 128, and a two-layer bidirectional gated recurrent unit (BiGRU) to model temporal dependency within the sequence of eight facial frames. An attention-pooling mechanism additionally highlights the most instructive temporal characteristics and yields a brief sequence-level representation for binary categorization. In parallel, a forensic decoder reconstructs 64 × 64 Laplacian residual maps, encouraging the network to learn edge-level and texture-level manipulation traces. The framework is trained end-to-end using a weighted combination of cross-entropy classification loss and L1 reconstruction loss. Experiments on the Celeb-DF dataset, comprising 1168 cropped face videos, achieved 93% training accuracy, 90% frame-level classification accuracy, and an ROC-AUC of 0.92. Complementary analysis across spatial, frequency, and temporal feature domains showed that spatial texture characteristics provided the clearest class separation, while the forensic reconstruction branch achieved an MSE of 0.010. Overall, the results demonstrate the value of jointly integrating spatial encoding, temporal sequence modelling, attention-based aggregation, and reconstruction-guided forensic learning for deepfake detection.