What AuthentiScan is
AuthentiScan estimates whether media may be AI-generated or manipulated. We combine our in-house trained ML image model with classical forensics (ELA/FFT), JPEG/EXIF cues, C2PA checks, external providers where available, and text analysis. The app presents an interpretable score with the top contributing signals.
Our approach
- In-house ML for images: trained on curated real vs. AI datasets and calibrated for probabilistic output.
- Media forensics: error-level analysis (ELA), frequency-domain texture (FFT), JPEG grid/cues, and EXIF/C2PA provenance checks.
- Ensemble fusion: each detector contributes a score and confidence; we blend these into a single probability and show the breakdown.
Roadmap
- Video: extend ML to temporal models alongside frame-level analysis.
- Text: add ML classifiers for long-form and short-form writing to complement heuristics.
- Mixed content: specialized models for web pages, documents, and embedded media.
- Broader training data: expand genres, devices, codecs, and post-processing coverage.
Principles
- Signal, not verdict. Outputs are probabilistic and should be combined with context.
- Transparency. We surface rationale, per-signal contributions, and provenance indicators.
- Privacy. Uploads are processed transiently for analysis.
Limitations
Recompression, resizing, re-uploads, and camera pipelines can shift signals. Content Credentials, when present, may provide strong provenance information. No detector is perfect—use results alongside human review.
Contact
Email: authentiscanai@gmail.com