Frequency Domain Analysis

The core mathematical operation that converts an image from its spatial representation (pixels) into the frequency domain, representing it as a sum of sinusoidal waves of varying magnitudes and phases. This translation exposes periodic structures as distinct, localized peaks in the frequency spectrum. In forensics, the DFT is essential for identifying grid-like artifacts from neural network upsampling layers, which manifest as high-energy spikes at specific frequencies corresponding to the kernel size and stride of transposed convolutions.

A technique that reduces the 2D Fourier magnitude spectrum to a 1D curve by averaging power across all orientations for each frequency radius. This analysis reveals the spectral decay signature of an image. Natural photographs exhibit a characteristic 1/f power-law decay, while AI-generated images often deviate, showing anomalous high-frequency energy due to the lack of physical diffraction limits in neural rendering. Comparing the azimuthal average of a suspect image against a known authentic camera model's profile is a robust method for global synthetic detection.

Generative Adversarial Networks (GANs) using transposed convolutions leave a distinct forensic trace: regular, high-magnitude peaks in the frequency spectrum. These peaks correspond to the upsampling factor of the generator. By applying a peak extraction algorithm to the DFT magnitude spectrum, analysts can isolate these unnatural frequencies. The pattern of these peaks acts as a unique architectural fingerprint, allowing attribution of a synthetic image to a specific GAN family or even a known model instance, independent of image content.

A spatial-frequency hybrid method where a high-pass filter is applied to isolate noise residuals and edge information. In a composite image, the boundary between a spliced region and the host background creates a sharp discontinuity in the noise pattern. By analyzing the high-frequency residual, forensic tools can detect inconsistent noise variance or mismatched sensor pattern noise at the splice boundary. This technique is particularly effective against copy-move forgeries where the duplicated region has a different compression history.

Unlike the DFT, which uses infinite sinusoidal basis functions, the Discrete Wavelet Transform (DWT) decomposes an image into localized wavelets, capturing both frequency and spatial location. This is critical for detecting localized tampering. By separating the image into approximation (low-frequency) and detail (high-frequency) sub-bands, analysts can identify anomalies in edge sharpness or texture consistency at specific coordinates. DWT is highly effective for detecting inpainting artifacts where a region has been seamlessly reconstructed.

A specific spectral pattern appearing as a series of equally spaced, parallel peaks in the 2D Fourier magnitude spectrum, resembling a comb. This artifact is a definitive indicator of nearest-neighbor upsampling or certain types of pixel-domain duplication. In deepfake detection, comb artifacts can arise from the resizing operations within an autoencoder's bottleneck or from the face-swapping process where a warped face region is pasted and resampled onto a target background, leaving a periodic resampling trace.