Phân tích tần số

Frequency analysis uses DCT (Discrete Cosine Transform) to decompose images into frequency components. AI images typically lack the high-frequency detail found in real photographs.

AI diffusion models generate images at lower resolutions and upsample them. This process smooths out fine details, reducing high-frequency content compared to sensor-captured photos.

Real photos typically have a high-frequency ratio above 0.20. AI-generated images often fall below 0.15, indicating their synthetic smooth nature.

We convert images to grayscale, apply 2D DCT, and separate the spectrum into low and high frequency regions. The ratio of energy in these regions indicates image origin.

Artificial sharpening can boost high frequencies, but the pattern differs from natural sensor noise. Our analysis considers both ratio and distribution characteristics.

Frequency analysis is highly reliable for detecting diffusion models but can be affected by image filters. The 20% weight reflects its strong but not absolute discriminative power.

DCT produces only real values (no complex numbers) and has better energy compaction for image signals. It's also what JPEG compression uses, making it well-suited for forensics.

AI upscalers add synthetic high-frequency detail, but these patterns are uniform and lack the natural variation of camera-captured fine detail. Detection remains effective.

We process images at 512x512 maximum for analysis. Higher resolution provides more frequency data, but the ratios remain consistent across sizes.

Frequency analysis is particularly effective against diffusion models (DALL-E, Midjourney, Stable Diffusion). Older GAN-based generators show different but still detectable patterns.