A Dual-Stream CNN-LSTM Framework with ELA Preprocessing for Deepfake Detection

The proliferation of deepfake content presents critical threats to digital authenticity, privacy, and information integrity. This paper proposes a hybrid deepfake detection framework that integrates Error Level Analysis (ELA) with the ResNet-50 convolutional neural network to identify manipulated images. The system employs a dual-stream architecture that processes both raw frames and ELA-transformed inputs, enhancing spatial and temporal feature extraction through Long Short-Term Memory (LSTM) layers. Evaluated on both general-purpose and celebrity-based datasets, the proposed model demonstrates superior classification accuracy and robust generalization, particularly in detecting forgeries involving high-profile individuals. Experimental results affirm the model’s effectiveness in capturing compression inconsistencies and subtle manipulations, positioning it as a scalable and reliable solution for deepfake detection in real-world scenarios.