AI-Generated Image Detection: An Empirical Study of Generalization and Training Stability

AuthentiNet-Scratch: From-Scratch Learning for AI-Generated Image Detection

AuthentiNet-Scratch is a deep learning research project focused on detecting AI-generated images and distinguishing them from natural photographs. The proposed approach uses an EfficientNet-B0 convolutional neural network trained from random initialization, without pretrained weights, on selected subsets of the GenImage dataset. The study explores model training stability, validation performance, generalization across different image-generation methods, and robustness to image transformations such as compression and resizing.

The framework incorporates systematic data preprocessing, model optimization, checkpoint selection, and evaluation using metrics such as accuracy, precision, recall, F1-score, and ROC-AUC. It also emphasizes generator-wise evaluation, confusion-matrix analysis, and data-leakage prevention to assess the reliability of the detector beyond familiar training data.

Preliminary project logs indicate validation accuracy of approximately 83.6%, which requires verification and further testing. Future experiments will focus on unseen-generator evaluation and robustness analysis to determine how effectively the model generalizes to different AI image-generation techniques.

The primary objective of AuthentiNet-Scratch is to establish a reproducible baseline for AI-generated image detection while highlighting the importance of transparent evaluation, generalization, and responsible interpretation of model predictions.