DEEP-FAKE DETECTION USING HYBRID NEURAL NETWORK ARCHITECTURES

Abstract

The rise of deep-fake technologies poses significant challenges to media integrity, political trust, and public security. Advances in generative adversarial networks (GANs) have made synthetic content increasingly realistic, necessitating more robust detection methods. This study proposes a hybrid neural network architecture combining Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN—specifically LSTMs), and autoencoders for effective deep-fake detection. Utilizing benchmark datasets such as FaceForensics++ and the DeepFake Detection Challenge, the data underwent resizing, normalization, and extensive augmentation. CNNs captured spatial features, RNNs encoded temporal inconsistencies, and autoencoders detected fine-grained anomalies via reconstruction loss. These components were integrated at the feature level into a unified model trained with the Adam optimizer and categorical cross-entropy loss, validated through stratified k-fold cross-validation. The hybrid model achieved superior performance with accuracy of 96.4%, precision of 95.8%, recall of 96.1%, F1-score of 96.0%, and ROC-AUC of 0.982, outperforming baseline models. Ablation studies confirmed the critical contribution of each component. This hybrid approach offers a promising, scalable solution for maintaining information authenticity in the digital era.