Comparative Analysis of Deepfake Detection Models

Deepfakes, which leverage advanced machine learning techniques such as generative adversarial networks (GANs), pose a significant threat to the integrity of visual content and raise concerns about misinformation and identity theft. This research provides a comparative analysis of various deepfake detection models, aiming to dissect their strengths and weaknesses. Notable architectures like Xception and ResNet50 exhibit high accuracy, precision, and recall with minimal gender bias. However, the Swin Transformer, while excelling in fake image detection, faces challenges with real images, suggesting potential bias. The CNN model demonstrates subpar performance, emphasizing limitations in classifying both fake and real images effectively. MobileNet shows moderate overall performance but maintains balanced precision and recall. The study recommends an ensemble approach to combine model strengths and address individual weaknesses. Future work should focus on refining model architectures, exploring ensemble strategies, and mitigating biases in real image detection.