Color Space Analysis for AI-Generated Image Detection on CIFAKE: CNN Models and Statistical Evaluation

Rapidly evolving generative artificial intelligence (AI) presents both opportunities and threats. Developing software that can distinguish AI-generated content from authentic content is crucial for mitigating these threats. This study systematically analyzes the impact of color-space representations on AI-generated image detection using the CIFAKE dataset and further investigates whether simple fusion strategies improve discriminative performance. In the first stage, seven raw color spaces (RGB, YCbCr, Lab, HSV, LCh, Wavelet, and Opponent) are evaluated across three lightweight backbones (EfficientNet-B0, RegNetY-002, and ReXNet-100). In the second stage, the color spaces that consistently perform best across backbones (YCbCr and Lab) are further examined with additional feature maps (gray level, edges, Laplacian, and noise) to form fusion variants. In total, we conduct 144 runs (3 backbones × 16 modes × 3 random seeds). An ensemble approach is employed by averaging predictions from three random seeds, and performance is supported using bootstrap-based confidence intervals as well as Friedman and FDR-corrected Wilcoxon statistical tests. The best configuration achieves 98.91% accuracy, 98.91% F1-score, and 99.93% AUC on CIFAKE. Overall, the results show that selecting suitable color-space representations (notably YCbCr and Lab) and their simple fusions can yield consistent gains on the CIFAKE benchmark even with low-complexity architectures.