Detection Methodology

When a camera captures a photograph and saves it as JPEG, the entire image is compressed simultaneously using a Discrete Cosine Transform (DCT) algorithm operating on 8×8 pixel blocks. Every region of that image shares the same compression generation, the same "quality history." AI-generated images and composites do not share this uniform history. When a diffusion model renders an image, it generates pixels through iterative denoising rather than optical capture. When an editor pastes an AI face onto a real photograph, the spliced region carries different compression artifacts than the original. ELA works by re-saving the image at a known quality level (typically 95%) and computing the absolute difference between the re-saved version and the original. In a pristine photograph, this difference map, the ELA map, is relatively uniform across all regions. In a manipulated or AI-generated image, regions with different compression histories light up with elevated error values, appearing bright in the ELA visualization against a darker background. This technique is particularly effective for detecting image compositing. A face with a dramatically different ELA response than the surrounding background is a strong signal of digital insertion. Pure AI-generated images also exhibit ELA patterns characteristic of the generator's output format, typically lacking the graduated compression artifacts that distinguish real photographic content. Important limitation: images that have been re-compressed multiple times (for example, saved through social media platforms which re-encode uploaded files) accumulate compression artifacts that can complicate ELA interpretation. This is why ELA is used as one signal among several rather than as a standalone verdict.

Every physical camera sensor, whether CMOS or CCD, contains manufacturing imperfections that cause individual pixels to respond slightly differently to the same amount of incident light. This pattern, called Photo Response Non-Uniformity (PRNU), is unique to each physical sensor and consistent across every photograph that sensor takes. Think of it as the camera's hardware fingerprint. Because PRNU is a fixed property of the physical sensor, it appears as a consistent, low-amplitude pattern superimposed on the image signal in every photograph taken with that device. Statistical analysis of a single image can detect whether the PRNU pattern is consistent with a real camera sensor, absent entirely, or inconsistent between image regions. AI-generated images have no PRNU. They are produced by software running matrix operations on a graphics card, a fundamentally different physical process that produces no sensor fingerprint. The noise that appears in AI images is either an approximation added by the model (which lacks the statistical properties of real PRNU) or absent altogether. Our analysis examines the noise residual, what remains after the signal content of the image is subtracted, and measures its statistical properties. Real photographs exhibit Gaussian noise distributions with properties consistent with physical sensor behavior. AI-generated images either exhibit no consistent noise floor, show repetitive patterns that betray a computational origin, or present a noise distribution inconsistent with any known camera hardware profile. The "too perfect" image (zero noise, immaculate skin, flawless gradients) is actually more suspicious forensically than a slightly grainy photograph, because it lacks the expected signatures of physical capture.

Human vision operates in pixel space - we see brightness and color at specific locations. Forensic analysis also works in frequency space, where images are represented as combinations of sine waves at different frequencies and orientations. This transformation, called a Fourier Transform (specifically a 2D Discrete Fourier Transform), reveals structural patterns in images that are completely invisible to the eye. Real photographs, captured through a camera lens on a physical sensor, have frequency spectra shaped by optics and quantum noise. Camera lenses act as natural low-pass filters, smoothly attenuating high-frequency components. The resulting frequency spectrum rolls off gracefully from low to high frequencies with no periodic structure. AI image generators, both GANs and diffusion models, produce images through learned convolutional operations and upsampling steps. These operations introduce characteristic periodic patterns into the frequency domain. Convolutional layers operating at fixed kernel sizes (typically 3×3 or 5×5 pixels) produce spectral peaks at corresponding frequencies. Bilinear or bicubic upsampling from lower-resolution intermediate representations introduces spectral artifacts at frequencies corresponding to the upsampling ratio. These spectral signatures are consistent across images generated by the same architecture and often survive significant image post-processing. In the frequency domain, an AI-generated image often displays a faint grid pattern or periodic spectral peaks that are absent in photographs of equivalent visual quality. This technique is particularly powerful for model attribution - identifying not just that an image is AI-generated, but which specific model family produced it, based on its spectral fingerprint. Research groups have demonstrated attribution accuracy above 90% for known model families using this approach.

Light behaves according to the laws of physics. Shadows cast by a single light source at a fixed position must all converge toward the same vanishing point behind the shadowed object. The angle of incident light must be consistent with the angle of the specular highlights on reflective surfaces. The intensity and color temperature of light must be consistent across objects in the same scene that are equidistant from the source. AI image generators violate these physical constraints regularly. Diffusion models like Midjourney, DALL-E 3, and Stable Diffusion are trained to produce aesthetically appealing images, not physically accurate simulations. They learn the statistics of how shadows and highlights appear in training photographs, but they do not model the underlying physics. The result is images where shadows look plausible locally but are geometrically inconsistent globally. Our shadow logic analysis works by identifying shadow edges and highlight regions in the image and computing the implied light source vector for each. For a real photograph, these vectors should all converge. For an AI-generated image, they frequently disagree - the shadow on a nose may imply a light source at 45° above-right, while the shadow on the same subject's collar implies a light source at 330° above-left. This inconsistency is not always obvious to the eye, because human visual perception is remarkably tolerant of local inconsistencies. We fill in lighting information contextually. The algorithm does not fill in - it measures. And the measurements frequently reveal impossible physics in AI-generated scenes. This technique is particularly robust to social media re-compression and format conversion, because it analyzes the semantic content of the image (where shadows fall) rather than low-level pixel statistics. Compression can destroy ELA evidence. It cannot straighten a shadow cast from the wrong direction.

Every photograph taken with a digital camera embeds a structured block of metadata, called EXIF (Exchangeable Image File Format) data, into the image file at the moment of capture. This metadata includes the camera manufacturer and model, lens specifications, exposure settings (ISO, aperture, shutter speed), GPS coordinates (if enabled), firmware version, and the precise timestamp of capture. AI-generated images lack real EXIF data. When a diffusion model outputs an image file, there is no camera that took the photo, no GPS module, no physical shutter, no real timestamp of a moment in physical space. The output file either has no EXIF data, contains generic placeholder values inserted by the generation software, or carries only a file creation timestamp from the user's operating system. Our metadata analysis examines three dimensions: First, presence and completeness: Does the file contain EXIF data at all? Are expected fields present (camera make, model, DateTimeOriginal, exposure settings)? Are they internally consistent? A file with a camera model field but no lens data, or a timestamp that predates the claimed event, raises immediate flags. Second, technical consistency: The exposure settings embedded in EXIF must produce the visual characteristics visible in the image. ISO speed determines noise level. Aperture determines depth of field. Shutter speed determines motion blur. An image with EXIF claiming ISO 100 in bright sunlight should have virtually no visible noise - if the noise level is inconsistent with the stated parameters, the metadata has been fabricated. Third, software chain analysis: The EXIF Software field records which application last processed the image. Some AI generation pipelines and compositing tools leave identifiable strings in this field. Cross-referencing the software chain against the claimed capture circumstances can reveal processing inconsistencies. Important caveat: social media platforms routinely strip EXIF data from uploaded images. Absence of EXIF is therefore inconclusive for images sourced from social media. Its presence and internal consistency, however, is meaningful positive evidence.

Generative Adversarial Networks and diffusion models, despite their different architectures, share a fundamental characteristic: they are trained processes that converge toward a specific statistical solution. That solution, the learned mapping from noise to images, is not perfectly random. It is shaped by the architecture's inductive biases, the training data distribution, the optimization algorithm, and the regularization choices made by the model's developers. These biases manifest as a consistent, low-amplitude statistical pattern in every image the model generates. This pattern, called a GAN fingerprint or model fingerprint, is analogous to the rifling marks on a bullet: invisible in normal use, but uniquely identifying under forensic examination. GAN fingerprints are extracted from the high-frequency residual of an image - what remains after the low-frequency signal content (the visible image) is subtracted. This residual contains the "fingerprint noise" characteristic of the generating model. Because this fingerprint is a property of the model's architecture and training, it is consistent across images regardless of their visual content. Our extraction process uses a bank of high-pass filters to isolate the fingerprint residual, followed by statistical characterization of the residual's spatial frequency distribution, spatial correlation patterns, and higher-order statistics. The resulting feature vector is compared against reference fingerprints from known model families. Research demonstrates that this approach achieves attribution accuracy above 90% for major GAN architectures (StyleGAN, ProGAN, BigGAN) and increasingly reliable attribution for diffusion models as reference databases expand. This means that for many images, we can identify not just that an image is AI-generated, but which specific model family (and sometimes which version) produced it.

No single forensic signal is infallible. ELA is sensitive to re-compression. PRNU is difficult to detect in highly compressed images. Shadow logic requires visible shadow regions to analyze. GAN fingerprints can be partially suppressed by adversarial post-processing. Any single signal can produce false positives or false negatives in isolation. The strength of multi-signal analysis is that it transforms independent, imperfect signals into a robust composite judgment. When five independent forensic signals all point in the same direction, the probability of a false verdict drops dramatically below the error rate of any individual signal. Our evidence fusion system operates as follows: Each forensic analysis module produces two outputs: a signal value (the measurement result) and a confidence weight (reflecting how reliable this signal is for this specific image - low for ELA on a heavily re-compressed file, high for shadow logic on a well-lit scene with visible shadows). These weighted signals are combined through a Bayesian evidence fusion model that accounts for the independence structure between signals. ELA and PRNU are partially correlated (both are sensitive to compression). Shadow logic and frequency analysis are largely independent. The fusion model uses this correlation structure to avoid double-counting evidence. The final output is a confidence score expressing how strongly the accumulated evidence supports the verdict. It is not a calibrated probability and should not be read as one: a 90% score does not mean nine in ten such images are AI-generated. Use it to rank how much weight to give a result, and read the evidence chain for the reasoning behind it. This calibration commitment shapes how we present results. We do not show binary verdicts ("FAKE" or "REAL"). We show probability estimates with the contributing evidence chain, so you can see which signals contributed to the verdict, how strongly, and what the known limitations of each signal are for this specific image. The human judgment remains in the loop. The algorithms provide evidence. You evaluate it.