Durable Content Credentials in Practice: How C2PA Manifests, Time-Stamped Seals, Invisible Watermarks and Perceptual Fingerprints Survive Everyday Image Processing
Content Credentials based on the C2PA specification bind a signed provenance manifest to a media file. Because the manifest lives in the file container, it is lost whenever an image is re-encoded, which motivates “durable” credentials that add an invisible watermark and a perceptual fingerprint as soft bindings. We measure both halves of this design with open, file-based experiments. First, we sign 120 images (JPEG, PNG, WebP) and pass them through 28 operations of widely used imaging tools and libraries. Outside of byte-for-byte copies, no operation yields a valid manifest: 91.0% of 3,120 outputs lose the manifest entirely, and the remaining 9.0% keep it in a state that fails validation, so routine processing such as a lossless jpegtran pass, an ImageMagick re-encode or an added metadata tag produces a credential that reads as tampered. Second, we build a layered benchmark with 150 registered and 50 never-registered images and 44 attacks, in which every layer is decoded from the same attacked file. The manifest identifies no attacked image (0 of 6,600). An ISCC Content-Code fingerprint identifies 72.3%, a classical DWT-DCT-SVD watermark 42.7%, TrustMark 81.7% and the Trufo watermark 88.1%. Paired with the fingerprint, the two learned watermarks reach 85.7% and 95.3%, with no wrong identifications and no false matches in 2,250 probes of unregistered images per layer. The fingerprint fails on geometric edits, TrustMark on crops and rotation, and the Trufo watermark on strong compression and large rotations, so the layers are complementary. Third, we evaluate content-aware detached seals: an RFC 3161 time-stamp over a record holding the file hash, a canonical pixel hash and an ISCC code. A plain file-hash seal matched none of the attacked copies; the content-aware seal matched 72.3% (15.9% at pixel level, including metadata edits that break C2PA), again without wrong or false matches, at 1.7 KB per token versus about 95 KB for an embedded manifest. We release code and data and derive guidance for verifiers, including identity-verification systems, on how to interpret missing and broken credentials.
Authors
- Daniel Wattenhofer
- Artus William Guenes (ORCID: https://orcid.org/0009-0003-5918-9266)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24231656
- Primary Topic
- Advanced Steganography and Watermarking Techniques
- Type
- preprint