Decision-Aid Framework for Face Authentication Detection Using ResNext50 and BiLSTM to Enhance Media Integrity

This paper presents a decision-aid framework for face authentication detection that integrates ResNext50 with Bidirectional Long Short-Term Memory (BiLSTM) networks to enhance media integrity and improve deepfake detection. Unlike conventional approaches focusing solely on spatial features, the proposed hybrid model incorporates temporal analysis, enabling the detection of subtle manipulations distributed across sequential video frames. ResNext50 is a robust spatial feature extractor from cropped facial regions, while BiLSTM captures bidirectional temporal dependencies, enriching the model’s contextual understanding and temporal sensitivity. The model was evaluated on three benchmark datasets—FaceForensics++, DFDC, and Celeb-DF—and demonstrated superior performance in classifying real versus manipulated facial content. The hybrid ResNext50 + BiLSTM model achieved a final accuracy of 96.11%, precision of 96.42%, recall of 96.68%, F1-score of 96.23%, and an AUC of 98.89%, outperforming the standalone ResNext50 and other state-of-the-art CNN-based approaches. These improvements reflect a notable reduction in false positives and negatives, critical for real-world face authentication systems. The proposed framework offers a scalable and robust solution for digital forensics, media verification, and cybersecurity applications, where maintaining trust in visual content is paramount. This study underscores the pivotal role of temporal modeling in improving the reliability of face manipulation detection. Future work will explore advanced fusion strategies, lightweight architectures for real-time deployment, and cross-dataset generalization to enhance adaptability in diverse, real-world scenarios.