Skip to content

Metadata.md

Latest commit

 

History

History
144 lines (124 loc) · 7.45 KB

Metadata.md

File metadata and controls

144 lines (124 loc) · 7.45 KB
Raw

Metadata & Content Credentials

Code: analysis/metadata_analysis.py — entry points analyze_metadata(path) and read_c2pa(path).

What it does

Reads what the file says about itself and looks for internal contradictions. It does not produce a score or a verdict: metadata is written by software, is trivially edited or removed, and a clean record says nothing about the pixels.

Sources read (the file is never re-encoded):

Source How
EXIF IFD0, Exif IFD, GPS IFD, Interop, IFD1 (thumbnail) Pillow getexif() / get_ifd()
EXIF thumbnail bytes piexif
XMP packet (JPEG APP1, PNG iTXt, WebP, TIFF tag 700) parsed with ElementTree; packets containing a DTD/entities are refused
PNG tEXt / iTXt / zTXt / eXIf chunks Pillow img.text, info["exif"]
C2PA manifest store c2pa-python Reader, offline (no remote manifest fetch, no OCSP)

Pillow's info fields such as jfif, dpi, progressive are container properties, not EXIF, and are not counted as EXIF.

Checks and how to read them

Levels: info = neutral fact, notice = weak indicator, warning = inconsistency found.

  • EXIF presence. JPEG/TIFF without EXIF → notice: stripped by software or messaging apps, or never written. Absence is not evidence either way. PNG, WebP, GIF without EXIF → info (normal for the format).
  • Software. EXIF Software, ProcessingSoftware, XMP CreatorTool, XMP History softwareAgent, PNG Software are matched against a list of known editors (Photoshop, GIMP, Affinity, Pixelmator, Snapseed, PicsArt, Paint.NET, Pixlr, Canva, Luminar, Facetune, Photopea, …) → notice "written by image editor". Raw developers (Lightroom, Camera Raw, darktable, RawTherapee, Capture One, DxO, …) → notice "processed with raw developer/editor" — normal for photographers. Anything else (e.g. phone firmware strings) → info. An editor tag says the file was saved by that program, not what was changed.
  • Dates. DateTime (last modified), DateTimeOriginal, DateTimeDigitized, with OffsetTime* (used when both sides of a comparison carry one) and SubSecTime*.
    • unparseable, before 1839 or in the future → warning (impossible date);
    • Digitized earlier than Original → warning;
    • Modified earlier than Original → warning;
    • Modified more than 2 s after Original → notice (file written again after capture: editing, raw development, re-export). Blank or all-zero dates are treated as "not set". A wrong camera clock produces the same patterns.
  • EXIF thumbnail. The embedded IFD1 thumbnail is compared with the main image downscaled to the thumbnail size. Both are blurred (σ = 1 px), one gain/offset is fitted, and the residual is measured per 8×8 thumbnail block. Warning "thumbnail differs from main image — image content changed after the thumbnail was written" when correlation < 0.90 or the worst block's residual is > 9× the median block's. Different aspect ratio → notice only (cropping, or cameras that letterbox thumbnails). The figure shows thumbnail | downscaled main image | residual.
  • Pixel dimensions. Image size vs EXIF PixelXDimension/YDimension (either orientation) → notice when they differ (resized or cropped after the EXIF was written).
  • MakerNote. Absent while Make is a maker whose cameras normally write one (Canon, Nikon, Sony, Fujifilm, Olympus/OM, Panasonic, Pentax/Ricoh, Leica, Apple) → notice: software that rewrites EXIF often drops it.
  • GPS. Coordinates are reported only when both latitude and longitude are present. A GPS block with only a version ID is reported as "no fix", never as (0, 0).
  • XMP. xmpMM:History events (action, softwareAgent, when, changed) → notice with a table; photoshop:DocumentAncestors → notice (content from other documents was placed in this one); xmpMM:DerivedFrom → info; Iptc4xmpExt:DigitalSourceType of trainedAlgorithmicMedia / compositeWithTrainedAlgorithmicMedia → warning (declared AI-generated).
  • PNG text. Keys written by image-generation front ends (parameters, prompt, workflow, Dream, sd-metadata, invokeai_metadata, …) → warning "likely AI-generated output".
  • C2PA Content Credentials. Claim generator, signer / issuer, signing time, actions, ingredients, validation state and failure codes.
    • hash mismatch codes (e.g. assertion.dataHash.mismatch) → warning: the file was changed after it was signed;
    • any other failure (except signingCredential.untrusted, which is expected because no trust list is configured) → warning;
    • otherwise info "signature and hashes intact; signer identity unconfirmed";
    • an action with a trainedAlgorithmicMedia digital source type → warning (declared AI-generated content).

If no notice or warning is raised, the result says which checks ran and "no inconsistency found at this sensitivity".

Output

  • metrics: format, width/height (px), EXIF tag count, XMP history event count, software entry count, C2PA manifest count, and when a thumbnail is compared: thumbnail correlation and worst-block residual (× median).
  • tables (only those with content): Camera, Software, Timestamps, GPS, XMP history, PNG text chunks, Full EXIF (stringified, binary values shown as <N bytes>), C2PA, C2PA actions.
  • images: thumbnail comparison strip (when an EXIF thumbnail exists).
  • details: the same data as plain JSON (exif, xmp, c2pa, thumbnail, png_text).

read_c2pa(path) returns a plain dict: present, validation_state, valid, claim_generator, title, signer, issuer, signing_time, actions, ai_generated, ai_indicators, ingredients, manifests, failures, errors.

Measured performance (thumbnail check)

Seeded benchmark, 8 photos (assets/sample images/sampleImg.jpeg + 7 scikit-image photos) × 40 trials:

Case Flagged
Genuine 160-px thumbnail (nearest/bilinear/bicubic/Lanczos/box, thumb q50–95, main re-saved q20–100), n = 320 0 %
Stale thumbnail, pasted patch 2 % of frame 98.1 %
… 5 % 99.4 %
… 10 % 98.8 %
… 20 % 99.1 %

Correlation alone (the previous check) caught only 7–78 % of these. The bundled sample image carries a real stale thumbnail (the object in the main image is missing from the thumbnail): correlation 0.953, worst block 11.4× → flagged. tests/test_metadata.py re-runs a 20-case subset (FPR ≤ 5 %, TPR ≥ 90 %).

The other checks are rule-based and have no threshold to calibrate; they are covered by fixture tests (camera EXIF, Photoshop/Lightroom/firmware Software tags, stale thumbnail, GPS version only, PNG text chunks, XMP history, date inconsistencies, time-zone offsets, C2PA present / absent / tampered).

Limitations

  • Metadata can be edited, copied from another file or stripped; an inconsistency can have an innocent cause (wrong camera clock, batch tools, messaging apps), and consistency proves nothing about the pixels.
  • The thumbnail check only works when an editor left the old thumbnail in place, and only sees edits large enough to survive downscaling (≈ 2 % of the frame and up). Global tone edits are not detected (gain/offset is fitted out).
  • The editor and maker-note lists are finite; an unknown editor is reported as "not a known editor".
  • C2PA is validated offline with no trust list: signer identity is never confirmed, remote manifests and certificate revocation are not checked. An absent manifest is the normal case.