Analysis and Accuracy Improvement of Deep Learning Models for Ai-Generated Image Detection

Abstract

The rapid development of generative artificial intelligence makes authentic photographs increasingly difficult to distinguish from AI-generated images. This study analyzes EfficientNet-B3 and improves the detector using a Convolutional Block Attention Module (CBAM). A balanced subset of 5,000 CIFAKE images was used, containing 2,500 REAL and 2,500 AI-generated images. After duplicate bytes were removed, 4,000 images were allocated to stratified five-fold cross-validation and 1,000 images were retained as a held-out test set. Images were resized to 300×300 pixels and processed using gray-world color normalization, contrast enhancement, and training augmentation. The baseline achieved 88.25±1.20% validation accuracy and an 88.28±1.14% F1-score, whereas the initial CBAM configuration achieved 87.52±0.90% and 87.49±0.94%. However, the five-checkpoint CBAM ensemble reached 89.00% test accuracy and an 88.98% F1-score, improvements of 0.80 and 0.85 percentage points over the baseline ensemble. Ablation showed that CBAM-last achieved an 88.32% validation F1-score. The findings indicate that appropriate attention placement and checkpoint ensembling can improve performance; external test sets are recommended to assess generalization.

The rapid development of generative artificial intelligence makes authentic photographs increasingly difficult to distinguish from AI-generated images. This study analyzes EfficientNet-B3 and improves the detector using a Convolutional Block Attention Module (CBAM). A balanced subset of 5,000 CIFAKE images was used, containing 2,500 REAL and 2,500 AI-generated images. After duplicate bytes were removed, 4,000 images were allocated to stratified five-fold cross-validation and 1,000 images were retained as a held-out test set. Images were resized to 300×300 pixels and processed using gray-world color normalization, contrast enhancement, and training augmentation. The baseline achieved 88.25±1.20% validation accuracy and an 88.28±1.14% F1-score, whereas the initial CBAM configuration achieved 87.52±0.90% and 87.49±0.94%. However, the five-checkpoint CBAM ensemble reached 89.00% test accuracy and an 88.98% F1-score, improvements of 0.80 and 0.85 percentage points over the baseline ensemble. Ablation showed that CBAM-last achieved an 88.32% validation F1-score. The findings indicate that appropriate attention placement and checkpoint ensembling can improve performance; external test sets are recommended to assess generalization.