Deep Learning Approach for Robust Deep Fake Detection using CNN-GRU Architecture

Abstract

In recent years, the rapid advancement of deepfake generation techniques has posed significant challenges for content authenticity and security. To address these concerns, the adoption of deep learning methodologies for deepfake detection has gained substantial traction due to their superior accuracy over traditional methods. Among the popular architectures, Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN), particularly Gated Recurrent Units (GRU), have demonstrated notable effectiveness. This study presents a CNNGRU-based model tailored for deepfake detection, incorporating a typical pipeline that includes data preprocessing, feature extraction, and classification. The model capitalizes on subtle artifacts introduced by GAN-based deepfake
generators, enabling robust detection despite the increasing realism of synthetic content. Experimental results reveal a high training accuracy of 98.93% over 10 epochs and a test accuracy of 81.97%, indicating strong generalization capabilities. Despite minor fluctuations, the validation accuracy shows an overall upward trend, affirming the model's learning efficiency and stability across training phases. The proposed CNN-GRU framework thus offers a promising and contemporary solution for detecting deepfakes in dynamic and evolving data environments.