A Performance Analysis to Detect Synthetic Images with Deep Learning-Based Image Classification

This study tackles the growing public concern about the spread of AI-generated images and their impact on public trust in the media and society. The ability to distinguish between true and fraudulent contents has grown increasingly difficult, resulting in the spread of misinformation and the possibility of disinformation operations. Deepfake technology, in particular, offers a huge risk since it generates incredibly realistic manufactured films and images. To address this issue, we suggest the creation of an AI-based system capable of distinguishing between actual and fraudulent photos. We use a deep learning-based convolutional neural network (CNN) architecture, which is well-known for its image classification skills. The research focuses on improving performance by optimising the CNN model and hyperparameters. A 120,000-image CIFAKE dataset is employed, with 60,000 AI-generated synthetic photos and 60,000 real-captured photographs. To test the efficiency of the trained model, performance evaluation metrics such as accuracy, precision, recall, and F1-score are used. With a high accuracy rate of 95.9%, the proposed method achieves remarkable outcomes in properly distinguishing between actual and AI-generated images. Furthermore, the model has high precision, recall, and F1-score values, suggesting its robustness in dealing with both types of images. The study provides light on the issues posed by AI-generated images’ rising sophistication and emphasises the necessity for automated solutions to tackle fraudulent content. This research contributes to sustaining public trust in media and limiting the harmful societal impacts of AI-generated disinformation by attaining a high level of accuracy in deepfake identification.