Technical Documentation
Detection Methodology
A transparent, scientific overview of the forensic pipeline used to estimate whether submitted media is synthetic or authentic.
Overview
The detection system operates as a multi-layer forensic pipeline. Each layer extracts independent signals from different domains (metadata, frequency, geometry, neural classification) and contributes to a final probability score via a weighted geometric mean. No single layer can dictate the outcome; they must converge.
The system does not claim to provide absolute certainty. Its output is a forensic probability estimate, and all intermediate evidence is exposed to the user for independent evaluation.
Image Analysis Pipeline
Metadata & Provenance Analysis
The system extracts EXIF/XMP headers from the image file. It searches for hardware sensor signatures (camera make, model, exposure parameters, GPS data), software editor fingerprints, and explicit AI provenance markers such as C2PA (Coalition for Content Provenance and Authenticity) digital assertions.
Signal classification:
- Strong EXIF (camera model, focal length, ISO) — leans toward authentic origin.
- AI generator metadata (Midjourney, DALL-E, Stable Diffusion in Software tag) — definitive AI provenance.
- Absent or minimal metadata — neutral (common in screenshots, social media, and many AI outputs).
Error Level Analysis (ELA)
The image is re-compressed at a fixed JPEG quality level. The pixel-wise absolute difference between the original and re-compressed version is computed and amplified. In an unmodified photograph, the compression error is uniform across the frame. In composited or AI-generated images, regions with different compression histories produce visible anomalies.
Mathematical basis:
The ELA score is the mean of the amplified absolute difference: score = mean(|original - recompressed| * scale). The detector includes a visual ELA heatmap that uses a JET colormap to render the spatial distribution of compression anomalies.
FFT Texture Periodicity
A Discrete Fourier Transform (DFT) is applied to the grayscale image. The frequency spectrum is analyzed for abnormally strong periodic peaks. AI upscaling algorithms and diffusion model decoders frequently introduce repeating grid-like patterns at specific spatial frequencies that are absent in optical photographs.
Known limitation:
Social media platforms (Facebook, Instagram, WhatsApp) apply aggressive JPEG re-compression using 8×8 DCT block encoding, which introduces artificial periodicity indistinguishable from AI upscaler artifacts. The system detects these "orphaned media" conditions and automatically neutralizes the FFT signal to prevent false positives.
Sensor Noise Fingerprint (PRNU)
Every physical camera sensor introduces a unique, low-level pattern of photon shot noise and fixed-pattern noise (Photo Response Non-Uniformity). This noise is measured by computing the variance of the high-frequency residual after applying a median filter. AI-generated images, being mathematically synthesized, exhibit abnormally smooth pixel neighborhoods in homogeneous regions.
Computation:
The image is divided into a grid of patches. For each patch, the difference between the original and median-filtered version is computed. The variance of these residuals across patches quantifies the "noisiness" of the image. Low variance indicates mathematical smoothness characteristic of synthetic content.
Physical Geometry Engine (3D Depth Maps)
The system employs a pre-trained monocular depth estimation model (Depth-Anything-V2) to reconstruct a 3D depth map from the 2D image. It then computes Sobel gradients on both the visual image and the depth map. Regions where depth boundaries sharply diverge from visual edges indicate hallucinated geometry — objects that appear flat in 3D but have sharp 2D edges, or vice versa.
Asymmetric scoring:
This engine operates as a penalty-only indicator. If physical inconsistencies are detected, the score rises sharply. If the 3D geometry is consistent, the score remains neutral (0.5), since high-quality AI generators can produce geometrically correct scenes. This prevents "good AI" from being falsely validated as authentic.
Vision Transformer Classifier (ViT)
For images that fall within the "Uncertain" zone (approximately 35–72% confidence), a secondary neural network classifier is invoked. This is a Vision Transformer (ViT) model fine-tuned on deepfake detection. It provides a learned, high-dimensional assessment complementary to the handcrafted heuristics above.
Conditional execution:
The classifier is not invoked for every image. It runs only in the gray zone, and its ability to override the heuristic score is governed by confidence thresholds. For social media (orphaned) images, the override threshold is set to 80% to avoid false activations on heavily compressed content.
Weighted Geometric Mean Fusion
All evidence signals are organized into independent buckets (Metadata, Compression, Visual, Semantic, Physical). Each bucket is computed as a weighted geometric mean of its constituent signals. The final score is a weighted geometric mean of these buckets and the model score.
Unlike a simple arithmetic average, a geometric mean ensures that a single extremely low or high outlier cannot dominate the result. All evidence dimensions must converge for a confident verdict.
Orphaned Media Profile:
When the system detects that an image has been stripped of metadata and resized to typical social media dimensions (e.g., Facebook, Instagram), it activates a specialized weight configuration. Compression and FFT heuristics are demoted to near-neutral because the platform's own processing destroys these forensic signals. The model confidence and 3D physics weights are elevated instead.
Text Analysis Pipeline
Text detection combines a pre-trained RoBERTa transformer model (for English semantic analysis) with statistical stylometry heuristics (paragraph variance, sentence burstiness, type-token ratio, and vocabulary predictability) applicable across English, Czech, and Slovak.
Important disclosure:
Text detection is inherently more ambiguous than image detection. Formal, professional, or template-based human writing (legal documents, technical manuals, formulaic marketing copy) may exhibit low perplexity and uniform sentence structures that closely resemble LLM output. Conversely, AI-generated text that has been manually edited or re-arranged by a human may not retain detectable signatures. The output should be interpreted as a probabilistic signal, not a definitive authorship judgment.
Known Limitations
Social media compression. Platforms like Facebook, Instagram, and WhatsApp strip metadata and aggressively re-compress images. This destroys most heuristic signals. The system compensates with weight shifting, but heavily compressed images with no metadata will often result in an "Uncertain" verdict by design.
High-quality AI with no provenance. State-of-the-art generators (Midjourney v6, FLUX, DALL-E 3) can produce images with correct 3D geometry, natural noise textures, and no metadata traces. When distributed via social media, these images may be forensically indistinguishable from real photographs using current methods. An "Uncertain" result in such cases is the mathematically honest answer.
Text analysis is a signal, not proof. LLM-generated text that has been manually rewritten, or human text that is naturally formulaic, can produce misleading scores. The text detector is designed as an advisory tool to flag suspicious patterns, not to serve as evidence of authorship.
Adversarial evasion. A determined adversary can modify an AI-generated image to bypass known detection heuristics (e.g., adding synthetic EXIF, re-encoding with camera-like quantization tables). No detection system is immune to targeted adversarial manipulation.
Commitment to Transparency
Every analysis returns the full breakdown of intermediate signals, evidence bucket scores, raw feature values, and explanatory notes. The ELA heatmap can be viewed directly on the detection result. All scoring logic is deterministic and reproducible. We believe that a detection system that refuses to explain itself cannot be trusted, and we design accordingly.
View Public Benchmark Results