With the increasing realism of AI-generated images, accurately differentiating these from camera-captured images is critical for media authenticity, digital forensics, and security. This paper presents a binary classification model leveraging the EfficientNet-B0 architecture to distinguish AI-generated images (from sources like DALL-E and Stable Diffusion) from real, camera-captured images. By employing transfer learning, the model is trained on a balanced dataset and optimized through data augmentation, cross-entropy loss, and the Adam optimizer. Insights from PCA and random forest classification further enhance the model’s accuracy and generalizability. This paper details the model’s architecture, data preprocessing, training process, and evaluation metrics while discussing implications and future applications in real-world scenarios.