-
-
Notifications
You must be signed in to change notification settings - Fork 0
Glossary
Technical terms used in Jura Trace, explained in plain language. Cross-references appear in italic. For the full scoring algorithm behind terms like Trust Score and Verdict, see Methodology.
aHash (Average Hash). A perceptual hashing algorithm that reduces an image to an 8×8 greyscale grid and compares average cell brightness. Fast and effective for near-duplicate detection. Computed alongside dHash and pHash to build a three-signal fingerprint per asset. See also Perceptual Hash.
Assertion. In C2PA terminology, a single structured claim inside a Manifest. Assertions record facts about a file: who created it, what training-mining policy applies, what rights are attached, and what the digital source type is. Each assertion is hashed and included in the cryptographic claim.
Authentic (Verdict). The verdict assigned when no forensic detectors flag significant anomalies and the deepfake ensemble falls below the Synthetic threshold. Does not guarantee authenticity; it means automated analysis found no evidence of manipulation within current detection limits.
C2PA (Coalition for Content Provenance and Authenticity). An open standard for embedding cryptographic provenance metadata in digital files. Manifests record who created or edited a file, when, and with what tools. Jura Trace can sign assets with C2PA manifests and verify existing manifests from signers including Adobe, BBC, Canon, Leica, and Nikon.
Claim. The cryptographically signed data structure at the centre of a C2PA Manifest. Binds together a set of Assertions, ingredient references, a timestamp, and the signer's certificate. Any post-signing modification invalidates the signature. Encoded as a COSE-Sign1 structure.
CLIP (Contrastive Language-Image Pre-training). A neural network trained to understand images and text jointly (OpenAI). Jura Trace uses the open-source OpenCLIP ViT-B/32 implementation for zero-shot AI/authentic classification. CLIP embeddings also serve as input features for the UnivFD probe. Optional: requires approximately 350 MB model download.
Composite. An image created by combining elements from multiple sources. When two independent regional detectors (segmented ELA, colour temperature) flag the same image, Jura Trace applies a composite-evidence cap, capping the trust score at 55% regardless of other signals.
Conformant Mode (signing). One of two C2PA signing modes in Jura Trace. Uses a certificate from a C2PA Trust List authority, so the resulting manifest is trusted by third-party C2PA verifiers such as Adobe's Content Authenticity web tool. Feature-flagged off in v1.0 and planned for a later release. See also Sovereign Mode.
Content Credentials. Adobe's branded name for the C2PA standard. Both terms refer to the same underlying technology. Jura Trace's signed files are compatible with any C2PA-compliant verifier regardless of which term is used.
Copy-Move Detection. A forensic technique that searches for duplicated regions within a single image. Clone-stamp and copy-paste manipulation leave regions identical to other parts of the same image. Jura Trace uses SIFT descriptors with RANSAC verification and DBSCAN clustering to identify clone regions.
COSE-Sign1. The cryptographic envelope format used to sign a C2PA Claim (RFC 8152, single-signature variant). Encodes the claim payload, the signing certificate, and the digital signature in one binary structure. Tampering with the payload invalidates the signature.
CRL (Certificate Revocation List). A periodically published list of revoked certificates. C2PA verifiers check a certificate against the issuing CA's CRL to confirm it was not revoked at the time of signing. Jura Trace supports CRL checking in Enhanced network mode. See also OCSP.
Deep Mode. An investigation mode that runs the full forensic pipeline, including all regional detectors (segmented ELA, colour temperature). More thorough than Standard mode. Typical analysis time: 30–60 seconds. See Methodology for the full detector list.
Deepfake. Synthetic or manipulated media created using AI. Originally referred to face-swapped video; now broadly covers AI-generated images, video, and audio. Jura Trace's detection ensemble combines GBM v4 with the UnivFD CLIP probe to produce a continuous confidence score. See also GBM and UnivFD.
dHash (Difference Hash). A perceptual hashing algorithm that compares adjacent pixel brightness differences rather than absolute values. More sensitive to structural changes (cropping, content removal) than aHash. Used alongside aHash and pHash in Jura Trace's fingerprinting system. See also Perceptual Hash.
Digital Source Type. An IPTC metadata field (Iptc4xmpExt:DigitalSourceType) declaring how content was created. Key values: digitalCapture (photographed), trainedAlgorithmicMedia (AI-generated), compositeSynthetic (synthesised from components). Jura Trace reads this field as part of EXIF anomaly detection and embeds it in the C2PA c2pa.training-mining assertion when signing. See also Training-Mining.
DWT-DCT-SVD. A frequency-domain invisible watermarking technique using three sequential transforms (Discrete Wavelet Transform, Discrete Cosine Transform, Singular Value Decomposition). Typically survives JPEG compression at quality 70+, proportional resizing, and modest cropping. Not available in this release.
ELA (Error Level Analysis). A forensic technique that re-compresses an image at a known quality level and measures the difference from the original. Regions with a different compression history appear as bright patches in the ELA map. Jura Trace offers both whole-image ELA and Segmented ELA for regional analysis.
EXIF (Exchangeable Image File Format). Metadata embedded in image files by cameras and editing software, including camera model, capture timestamp, GPS coordinates, and a vendor-specific MakerNote block. Jura Trace evaluates twelve EXIF anomaly rules, five injection-detection sub-checks, and two XMP AI-provenance sub-checks as part of its EXIF Anomaly detector.
Fingerprint. A compact numerical representation of an image's visual content, computed using perceptual hashing. Jura Trace stores three variants (aHash, dHash, pHash) per asset for near-duplicate matching even after resizing, compression, or light editing. See also Perceptual Hash and Hamming Distance.
GBM (Gradient Boosted Machine). The machine learning classifier at the centre of Jura Trace's AI detection ensemble. GBM v4 is trained on an 84-feature vector from over 10,000 images across 14 generator families. Performance: AUC-ROC 0.9868, authentic false-positive rate 4.54%. Combined with UnivFD v10onnx into an ensemble score.
Hamming Distance. A measure of difference between two binary strings, counting positions where they differ. Two images with a Hamming distance of zero have identical hashes; a distance of 5 or below typically indicates near-duplicates. Jura Trace reports Hamming distances for aHash, dHash, and pHash comparisons.
Inconclusive (Verdict). The verdict assigned when forensic signals are mixed or insufficient for a confident classification. The most common outcome for legitimate photographs that have been processed or shared. It means the analysis cannot confidently classify the content, not that it suspects manipulation.
Ingredient. In C2PA terminology, a file used as input to the current file's creation. Ingredients are recorded in the Manifest with a content hash. When an ingredient itself carries a manifest, the relationship is recorded as parentOf, establishing a multi-generation provenance chain.
IPTC. The International Press Telecommunications Council standard for news metadata. Fields include caption, credit, copyright, and AI-provenance values such as DigitalSourceType. Jura Trace reads IPTC/XMP fields as part of EXIF anomaly detection. See also XMP.
JPEG Ghost. A forensic technique that detects double JPEG compression. When a region from one JPEG is pasted into another and re-saved, the spliced region retains artefacts from its original compression quality level. Comparing against a sweep of re-compression qualities reveals these hidden strata. Carries a 0.5× weight in the trust score pending validation against a real-world splice benchmark.
JUMBF (JPEG Universal Metadata Box Format). The binary container format used to embed C2PA manifest data inside JPEG, PNG, and other image files (ISO/IEC 19566-5). C2PA uses JUMBF to package the manifest store inside the target file without modifying image pixels.
LBP (Local Binary Pattern). A texture descriptor encoding each pixel's relationship to its neighbours as a binary string. Natural photographs have characteristic LBP distributions; AI-generated images often produce measurably different ones. LBP block variance is one of the highest-weighted features in Jura Trace's GBM classifier.
MakerNote. A vendor-specific EXIF block written by a camera containing proprietary data (lens corrections, scene modes, firmware state). Its presence or absence is a forensic signal. Cameras from mandatory-MakerNote vendors (Canon, Nikon, Sony, Fujifilm, and others) always write this block; its absence in a file claiming to come from one of these cameras is flagged as a potential injection indicator.
Manifest. In C2PA terminology, the cryptographic record embedded in a file describing its provenance: who created it, when, and with what tools. A manifest is cryptographically signed; any subsequent modification to the file invalidates the signature. Jura Trace displays manifest details in the C2PA verification panel.
Noise Analysis. Examination of the grain pattern produced by a camera sensor. Natural photographs have a consistent noise signature. AI-generated images often exhibit unnaturally smooth regions or noise patterns inconsistent with real sensor characteristics. Jura Trace maps per-block noise variance and flags anomalous blocks.
NPR (Neighbouring Pixel Relationships). A forensic technique that analyses statistical relationships between adjacent pixels. Natural photographs have characteristic correlation patterns that AI generation tends to disrupt. An on-demand investigation tool; does not contribute to the trust score.
OCSP (Online Certificate Status Protocol). A real-time protocol for checking whether a specific certificate has been revoked (RFC 6960). Unlike CRL, OCSP queries for the status of a single certificate without downloading an entire list. Jura Trace supports OCSP in Enhanced network mode as part of C2PA manifest verification.
Ollama. A local large language model runtime used optionally by Jura Trace. Ollama-backed features are gated off in v1.0 and planned for v1.0.1. When enabled, all inference runs on your device; no data leaves your machine.
parentOf. A C2PA relationship type indicating the current file was derived from one or more Ingredient files. When Jura Trace re-signs an image that already carries a C2PA manifest, the existing manifest is preserved as a parentOf ingredient, maintaining the full provenance chain.
Perceptual Hash (pHash). A fingerprinting technique that produces a compact numerical representation of an image's visual content. Unlike cryptographic hashes, perceptual hashes remain similar when an image is resized, compressed, or lightly edited. Jura Trace computes three variants per asset (aHash, dHash, pHash) for near-duplicate matching.
Provenance. The documented history of a piece of content: who created it, how it was modified, and through which hands it passed. C2PA provenance manifests are Jura Trace's primary provenance mechanism, providing a tamper-evident chain from original capture through every subsequent edit.
RAG (Retrieval-Augmented Generation). A technique that combines a large language model with a searchable knowledge base to verify factual claims. Not available in this release, pending corpus expansion and formal accuracy evaluation.
Segmented ELA. A regional variant of Error Level Analysis that divides the image into an 8×8 grid and analyses each cell independently. Localised manipulation that whole-image ELA might average out becomes visible in the grid. One of the regional detectors that contributes to the composite-evidence cap in Deep mode.
Shadow Consistency. A forensic technique that analyses light direction across image regions. In a genuine photograph, shadows and highlights point toward the same source. Composite images often contain elements from different sources, so shadows in different regions can point in conflicting directions. An on-demand investigation tool; does not contribute to the trust score.
Signing Mode. The C2PA certificate type used when Jura Trace signs a file: Sovereign Mode (self-signed, offline-first, available now) or Conformant Mode (Trust List authority certificate, interoperable with third-party verifiers, feature-flagged off in v1.0 and planned for a later release). See those entries for details.
Sovereign Mode (signing). One of two C2PA signing modes in Jura Trace. Sovereign mode uses a certificate generated on your device, not issued by an external authority. Manifests are cryptographically valid and tamper-evident but are not trusted by third-party C2PA verifiers. Suited to internal records and offline-first institutional archives. Available in the Community (free) tier. See also Conformant Mode.
Splice Boundary. A forensic technique that examines image edges for three manipulation signals: JPEG DCT grid discontinuities, noise level changes at boundaries, and feathering artefacts from soft-selection pasting. An on-demand investigation tool; does not contribute to the trust score.
Standard Mode. The default investigation mode. Runs EXIF anomaly analysis, C2PA manifest verification, whole-image ELA, noise analysis, copy-move detection, and the deepfake ensemble. Typically completes in 10–20 seconds. Use Deep mode to add regional detectors for composite image investigation.
Synthetic (Verdict). The verdict assigned when multiple forensic detectors flag strong AI generation or manipulation signals and the deepfake ensemble score exceeds the Synthetic threshold. Indicates high confidence that the content was artificially generated or significantly altered. A Synthetic verdict does not identify the specific tool used.
Training-Mining. A C2PA assertion (c2pa.training-mining) declaring whether the file's creator permits its use for AI training or data mining. Jura Trace derives this policy from the declared licence when signing. The assertion expresses intent for rights-aware crawlers; it does not technically prevent training.
Trust List. A directory of certificate authorities whose certificates are accepted as proof of legitimate C2PA signers, maintained by the C2PA Secretariat. Jura Trace validates trust-listed manifests in Enhanced network mode. Sovereign-mode certificates are not on the trust list. See also Conformant Mode and Sovereign Mode.
Trust Score. A percentage from 0% to 100% indicating how strongly the combined forensic signals support authentic content. Built from five components applied in sequence: forensic signal analysis (primary), EXIF consistency (corroborating, capped at 20%), a C2PA adjustment, a composite-evidence cap at 55%, and a deepfake verdict ceiling. A forensic confidence indicator, not a guarantee of authenticity. See Methodology for the full algorithm.
TSA (Time Stamp Authority). An independent service that countersigns a document with a trusted timestamp, certifying it existed at a specific moment. C2PA manifests in Jura Trace are countersigned by a TSA at signing time, so the provenance record remains verifiable even if the signing certificate later expires or is revoked.
UnivFD. A zero-shot AI image detector based on CLIP ViT-B/32 embeddings (Ojha et al. 2023). Jura Trace ships UnivFD v10onnx: a LogisticRegression classifier trained on over 50,000 samples with platform-forwarded and multi-format augmentation. Performance: AUC-ROC 0.9929, authentic false-positive rate 3.87%, AI recall 95.77%. Combined with GBM v4 into an ensemble score.
Verdict. The three-way classification assigned to verified content: Authentic, Inconclusive, or Synthetic. Determined by the deepfake detection ensemble score, not the trust score alone. The verdict and the trust score can diverge: an image with clean EXIF but strong AI generation signals will receive a Synthetic verdict. Consider both values together.
Watermark (Invisible). A signal embedded in the frequency domain using the DWT-DCT-SVD technique. Carries a short payload identifying the asset and the protecting institution, providing a provenance layer independent of file metadata. Not available in this release.
Weather Context. An optional corroboration lookup on the Verify panel. When an image carries GPS coordinates and a capture timestamp, retrieves historical weather conditions for that location and date from the Open-Meteo archive. Informational only; does not affect the trust score. Requires Enhanced network mode and makes one outbound request to Open-Meteo.
XMP (Extensible Metadata Platform). An Adobe XML-based metadata standard embedded alongside EXIF data. XMP carries IPTC fields, edit history (xmpMM:History), rights statements, and AI-provenance declarations (Iptc4xmpExt:DigitalSourceType, xmp:CreatorTool). Jura Trace parses XMP as part of EXIF anomaly detection. See also EXIF and IPTC.
Last updated 2026-06-04 for Jura Trace v1.0.