Authentica adds three-layer image checks for C2PA, watermarks, forensics
An article on Medium details the release and technical architecture of 'authentica,' a new open-source Python library designed to detect AI-generated and manipulated images. The library's approach combines three distinct methods: parsing C2PA (Coalition for Content Provenance and Authenticity) cryptographic manifests for provenance, detecting invisible watermarks through frequency-domain analysis (DCT, DWT, FFT), and performing image forensics to find statistical anomalies (ELA, noise residual analysis). The author provides a detailed breakdown of each layer, including Python code for parsing C2PA's JUMBF binary containers and verifying COSE signatures.
Key Takeaways
- C2PA manifests are embedded in JPEG APP11 segments and PNG caBX, iTXt, or tEXt chunks, and authentica reassembles split APP11 segments before parsing.
- The library verifies COSE_Sign1 signatures and checks embedded certificate chains, while also flagging mismatches between stored and computed SHA-256 hashes.
- Watermark detection combines DCT, DWT, and FFT scores, with detection triggered at a 0.55 confidence threshold and weighted as DCT 40%, DWT 35%, FFT 25%.
- Forensics uses error level analysis, noise residual analysis, and frequency artifact detection, then rolls those into a trust score with a 50-point neutral baseline.
- The author says the tested PNG scan took 2.47 seconds on a local Intel i5 machine with no GPU, and a typical 800×600 JPEG scans in under 300 ms.
Why It Matters
The immediate value here is operational: authentica packages provenance, watermarking, and forensic checks into one local Python tool rather than forcing teams to stitch together separate systems. That matters because the post argues C2PA alone is incomplete when most web images have no manifest. The broader angle is workflow integration: the library outputs ExifTool-style tags and offers CLI, PyPI, GitHub, and FastAPI examples, which makes it easier to slot into moderation or asset pipelines. What to watch is the v0.2.2 roadmap, especially the planned `authentica.video` module for frame-level C2PA and watermark analysis for MP4/MOV.
Read full article at medium.com