Forensic Truthfor Every Exhibit

Exiphore is a forensic media authentication platform built for law enforcement and security agencies. It examines video, imagery, audio, documents and text for signs of deepfake generation or manipulation, and produces an explainable, court-ready record: the measurements taken, their limitations, a sealed chain of custody, and a named examiner's signed adjudication.

Measured end-to-end through the deployed pipeline on a balanced 60-image sample of the public Hemg/deepfake-and-real-images corpus, the platform flagged 86.7% of manipulated exhibits with a 0.0% false-alarm rate, and returned undetected manipulations as inconclusive rather than wrongly cleared. The learned classifier scores ROC-AUC 0.983 through that same preprocessing. That is a small sample from one corpus of still images: it is a floor on discrimination, not a performance guarantee for casework. Accuracy falls substantially on compressed, re-encoded and redistributed media, the corpus under-represents South Asian faces, and learned detectors generalise poorly to generators absent from their training data. The platform lowers its own confidence accordingly rather than overstating a result, and the full conditions and limits are published on the site.

Six things, for every exhibit: the examination result (manipulated, authentic or inconclusive); a breakdown of which families of forensic evidence contributed and by how much; a confidence interval together with the quality ceiling the exhibit itself imposes; the sealed chain of custody and hash-chained audit trail; a named examiner's review, reasoning and signature; and a digest-sealed, court-ready report containing all of it, including the stated limitations.

Exiphore produces the record a court needs — measurements, stated limitations, a tamper-evident chain of custody, and a signed examiner adjudication — but an automated result is never an expert opinion on its own. Every finding requires a qualified examiner to review the measurements and record their reasoning before it is tendered. The platform is explicit about what it cannot establish.

Video, images, audio, documents and text. Each medium is examined by the measurements appropriate to it, and exhibits can be submitted individually, in batches, by URL, or from a monitored path.

No. Exiphore is deployed on your own infrastructure and evidence never leaves your estate. This is a requirement for data-sovereignty obligations and it is what makes the platform usable on material that cannot lawfully be sent to a third-party service. Optional narrative generation can be disabled entirely.

A single detector gives you one number with no way to interrogate it, and generalises poorly to generation methods it has not seen. Exiphore fuses eight independent families of forensic evidence, de-correlates related measurements so they are not counted twice, reports a confidence interval rather than a point estimate, and surfaces disagreement between signals instead of averaging it away.

The platform returns inconclusive and says why — which families disagreed, or what about the exhibit's quality limited the examination. It also recommends the investigative steps that do not depend on the detection result, such as preserving the source account and obtaining a higher-quality original.

Security agencies, law-enforcement agencies and police units handling synthetic-media offences: cyber crime, economic offences, women's safety, intelligence and forensic science laboratories. Access is by administrator-issued account only; there is no public sign-up.

Yes, and more precisely it is a forensic examination platform built around deepfake detection. A conventional deepfake detector returns one probability from one model. Exiphore runs measurements across eight independent families of forensic evidence, pools them so correlated signals are not counted twice, reports a confidence interval capped by the exhibit's own quality, and hands the result to a named examiner who reviews and signs it. Full detail is on the deepfake detector page.

In practice they describe the same category. Deepfake usually implies a manipulated human face or voice; synthetic media is the broader term, covering fully generated video with no real subject, cloned speech, AI-generated imagery, fabricated documents and synthetic identities. Exiphore examines all of them and reports which family of evidence supported the finding in each case.

In part. Learned classifiers only recognise what resembles their training data, so a generator released after the model was trained can evade that one signal entirely — the documented weakness of the whole detector family. This is why Exiphore never relies on a learned classifier alone: physiological, frequency, temporal and provenance evidence do not depend on having seen a particular generator before, so a new method has to defeat all of them at once.

Yes. Fingerprints derived from every exhibit are indexed, so the platform surfaces matches across cases — the same source media, the same production characteristics, or a recurring subject. These are presented as investigative leads requiring corroboration, never as identifications.

As a self-contained on-premises stack. It runs on standard server hardware without specialised accelerators, and can operate fully air-gapped once installed. Deployment, examiner training and validation against your own casework are part of onboarding.