Enhance and Implementation of Deep CNN-BILSTM Model for Deep Fake Image and Video Identification using Attention Mechanism

Abstract

With the rapid advancement of deepfake technology, ensuring digital media authenticity has become a critical challenge in today’s digitally connected world. Deepfake images & videos are increasingly being used for misinformation, identity manipulation, financial fraud, and reputational damage, making robust detection mechanisms essential. Traditional machine learning approaches often struggle to identify subtle visual and temporal inconsistencies present in highly realistic synthetic media. This paper suggests a hybrid deep learning architecture that combines a Deep Convolutional Neural Network (DeepCNN), Bidirectional Long Short-Term Memory (BiLSTM) & an attention mechanism for efficient deepfake detection in order to overcome these drawbacks. High-level spatial characteristics and fine-grained visual artifacts within individual frames are extracted by the Deep CNN component. The BiLSTM network captures temporal dependencies and motion-based inconsistencies across sequential frames, enabling the model to analyze dynamic patterns in video content. Furthermore, the attention mechanism selectively emphasizes the most discriminative spatial–temporal features, improving overall classification performance and reducing the impact of irrelevant information. The suggested model obtains an accuracy of 98.33% after extensive experimental assessment, demonstrating its robustness, dependability, and appropriateness for practical deepfake detection applications.