OPTIMIZING CNN PERFORMANCE FOR AI-GENERATED IMAGE CLASSIFICATION: A COMPARATIVE STUDY OF ARCHITECTURES AND OPTIMIZERS USING K-FOLD CROSS-VALIDATION

This study investigates CNN optimization for classifying AI-generated images. Using the CIFAKE dataset (60,000 real and 60,000 AI-generated images), we evaluated four CNN configurations with varying complexity and four optimization algorithms through 5-fold cross-validation. Our findings show Configuration 4 (4 Conv, 2 MaxPool) with Adam optimizer achieved the highest validation accuracy (0.8368±0.0135). Adam demonstrated consistent performance across architectures, while SGD showed strong but variable results improving with model complexity. Adagrad and Adadelta consistently underperformed. The final model achieved 85.28% test accuracy with balanced precision (0.8531) and recall (0.8528). Results indicate more complex architectures combined with adaptive optimizers like Adam provide superior performance for AI-generated image classification, with the balance between model complexity and optimizer selection being crucial. The consistent performance across real and fake categories demonstrates this approach's robustness for deepfake detection applications.