Performance Analysis of Deep Learning Architectures in Classifying Fake and Real Images

The advancements in artificial intelligence (AI) have significantly enhanced image manipulation capabilities, yet they also raise concerns regarding the proliferation of synthetic images. This study investigates the impact of Dynamic Dropout in optimizing deep learning models, including ResNet-101, DenseNet-201, VGG-19, and AlexNet, for classifying real and synthetic images using the CIFAKE and Real and Fake Face datasets. Dynamic Dropout was applied with a progressively increasing rate from 20 percent to 50 percent to enhance training stability and generalization. The results indicate that the optimal configuration consisting of 15 epochs, the Adam optimizer, and Dynamic Dropout consistently outperformed Static Dropout across all models. DenseNet-201 with Dynamic Dropout achieved the highest accuracy of 97.42%, with a precision of 97.33%, recall of 97.58%, and an F1-score of 97.45%. ResNet-101 and VGG-19 exhibited enhanced training stability, while AlexNet proved efficient for lightweight datasets. The Adam optimizer outperformed Nadam, offering greater stability in deeper architectures. Additionally, the 15th epoch was identified as the optimal training duration, balancing accuracy and overfitting mitigation. These findings underscore the importance of selecting optimal training configurations to enhance deep learning performance. Future research should explore adaptive dropout strategies, assess scalability on diverse datasets, and validate these techniques in real-world applications such as digital forensics and AI-generated content detection