Statistical Pattern Analysis

Entropy measures information content based on pixel value distribution. Natural images have specific entropy ranges; AI images often have abnormally high or low entropy.

Benford's Law describes the expected frequency of leading digits in natural data. Real DCT coefficients follow this distribution; AI-generated images often deviate from it.

Analyzing the distribution of pixel values. Real photos have smooth histograms shaped by scene content; AI images may have gaps, spikes, or unusual symmetry.

This measures how pixel values correlate with neighbors. Real images from sensors have specific correlation patterns; AI generation creates different spatial relationships.

AI generates images through mathematical processes that create detectable patterns. Even when visually perfect, statistical fingerprints remain different from camera output.

Editing affects statistics, which is why this method has lower weight. Heavy editing can normalize AI statistics, but typically introduces other detectable artifacts.

Benford's Law predicts ~30% of first digits should be "1", ~17% "2", etc. AI DCT coefficients often show flatter distributions not matching these natural ratios.

Statistical analysis alone has 70-80% accuracy. It works best combined with other methods, providing additional signal especially when visual analysis is uncertain.

Both global statistics and local patch statistics are computed. AI images often show suspicious uniformity of statistics across patches that would naturally vary.

Higher resolution provides more data for reliable statistics. Very small images may not have enough samples for accurate Benford analysis; minimum 256x256 is recommended.