The Hybrid Residual–BiLSTM Model for Mitigating the Vanishing Gradient Problem in Synthetic Image Detection

Abstract

The rapid advancement of generative artificial intelligence has enabled the creation of highly realistic synthetic images, raising concerns regarding the authenticity of digital media. While convolutional neural network (CNN)-based methods have shown promising results for synthetic image detection, deep architectures often suffer from training instability caused by the vanishing gradient problem and limited representation learning. This study proposes a Deep Residual BiLSTM Fusion model that integrates residual-based CNN feature extraction with bidirectional long short-term memory (BiLSTM) modeling to enhance training stability and contextual representation learning. The proposed model is evaluated on the CIFAKE dataset, a challenging low-resolution benchmark for synthetic image detection. Experimental results demonstrate competitive performance with an accuracy of 97.27% and balanced precision and recall across both real and synthetic classes. Gradient flow analysis further confirms stable optimization behavior, highlighting the effectiveness of the proposed hybrid architecture for robust synthetic image detection.