Detecting synthetic or altered media is no longer something manual review can keep up with. The volume and sophistication of fabricated images call for a layered, programmatic approach that goes beyond what a reviewer can catch by eye. Metadata checks and visual spot-checks were enough when fakes were rare and crude. Generative AI changed that, and prfd combines forensic analysis with cryptographic sealing to answer the question a human reviewer can no longer answer on their own.
Why human review stops working
An AI image generator does not capture light bouncing off a physical object. It constructs a canvas pixel by pixel from statistical probabilities. The result can look completely natural to a customer service rep or a claims adjuster, even though it is structurally different from a photo a camera sensor actually took. The signal that gives a generated image away is not something a person can see by looking harder. It has to be measured.
Four forensic layers, then the seal
When an application sends an image to the prfd endpoint, the system looks past what the photo shows and inspects the file itself. Verification runs across four layers:
- Metadata extraction. The system reads EXIF data, embedded structural signatures, and compression history, looking for the anomalies that tend to show up in a generated or edited file: a missing camera make, a rewritten segment, an editor's own tag.
- Pixel-frequency analysis. A spectral pass looks for the repeating mathematical artifacts generative models leave behind, patterns that stay invisible to a human reviewer no matter how closely they look.
- Sensor and provenance forensics. The system measures the noise floor a physical camera sensor produces, localizes regions spliced in from another image, and checks the file's compression statistics against Benford's law, a distribution real photographs follow and manipulated ones tend to drift from. When a file carries C2PA Content Credentials, they're verified cryptographically.
- Cryptographic sealing. Once a file clears the checks above, prfd binds the photo, its metadata, the time, and the verdict into a single signature, a seal anyone can verify and nobody can quietly edit afterward.
What the seal actually proves
Catching a fake is only half of it. The other half is making the files that are genuine provably so. A seal is a tamper-evident record bound to the file at the moment of capture. It does not just say the photo passed a check, it guarantees that nothing about the file, the pixels, the metadata, the verdict, changed between the moment it was sealed and the moment someone checks it later. That is the difference between a fraud score sitting in a database, which anyone with write access can edit, and a seal that breaks the instant anything underneath it changes.
Dropping into what you already have
None of this requires a team to build and maintain their own detection models. Integration is one REST call from the upload handler already in place: sign on capture, verify on read. The checks run on our infrastructure, and a client application sees a verdict come back inline with the upload, not a new system to operate.