Examination result
Manipulated, authentic or inconclusive — with the interval that bounds it.
Synthetic media forensics
Forensic examination of video, images, voice, text and documents — with measurements, stated limits and a signed report.
Interval 0.59–0.88 · Quality ceiling 57% · 13 of 14 signals used
A circulated video of a serving officer, Jantar Mantar, New Delhi — examined October 2026. Exhibit redacted; figures from the published case study.
The deliverable
A finding on its own is a number. What leaves the platform is a record an examiner can sign and another expert can interrogate.
Manipulated, authentic or inconclusive — with the interval that bounds it.
Which families of evidence contributed to the finding, and by how much.
How far the exhibit itself allowed the examination to go.
Hash, timestamps and custody events, each linked to the one before.
A named examiner's reasoning, recorded before release.
The complete record, downloadable and tamper-evident.
A face and a voice can be generated in minutes, well enough to pass a video call and well enough to reach a case file. Investigators are being handed media they cannot vouch for, and defences are claiming the same of media that is genuine.

deepfakes estimated to be shared in 2025, up from 500,000 in 2023.3
of organisations hit by a deepfake attack in the past twelve months.4
job candidates worldwide projected to be fake by 2028.5
The platform
One examination record for every kind of exhibit — each signal reported with what it can, and cannot, establish.
Explore the platform
Identity redactedExhibit · circulated video of a serving officer, identity redacted
Light reflected from skin changes with every heartbeat. Exiphore recovers that pulse separately from four regions of the same face — forehead, both cheeks, chin. On a real face they agree, because one heart drives all four. On a synthesised face they do not.
Pulse recovered independently from four facial regions disagreed: 98, 61, 48 and 100 bpm. A face perfused by one heart does not disagree with itself.
What it cannot tell youPulse recovery needs a face large enough and stable enough to measure. Below roughly 100 pixels, or under heavy compression, the signal degrades and the examination says so rather than guessing.

Example · a portrait whose right half is reconstructed, not captured
Natural images follow a power law: energy falls away smoothly as spatial frequency rises. Upsampling stages inside a generator do not reproduce that tail — they leave either too little high-frequency energy or a periodic pattern where noise should be.
The measured spectrum departs from the fitted natural-image power law across the high-frequency band — the shaded region — which is where a generator's upsampling stage shows itself.
What it cannot tell youRe-compression destroys this evidence. A screenshot of a screenshot can be genuine and still fail the test, which is why frequency never decides a finding on its own.

Example · spectrogram of a synthesised voice
Neural vocoders reconstruct very little genuine energy above about 7–8 kHz, and they produce a noise floor that is flatter than any room. A human larynx also never repeats a pitch period exactly; synthesis often does.
Energy falls off a cliff at the vocoder's reconstruction ceiling, and the floor above it is spectrally flat — room tone is coloured, synthesis is not.
What it cannot tell youA narrow-band phone call is band-limited too, for entirely innocent reasons. The examination weighs the cliff against the codec the file actually used.
Example · keyed-token test on a generated paragraph
Some generators embed a key-seeded watermark: a vocabulary split at every position into two sets, with watermarked output over-selecting one. Detection is a one-proportion z-test on that fraction, and it needs only the issuer's key — never the model that wrote the text. Most generators embed no watermark at all, and for those the text is reported as unexaminable.
The keyed-token fraction sits well above the 0.5 expected by chance, with a z-score past the decision threshold. Alongside it, sentence-length burstiness sits in the narrow band typical of model output rather than human writing.
What it cannot tell youA watermark hit shows the text came from a watermarked generator. It does not show who prompted it, and its absence proves nothing — most generators ship no watermark at all, so unwatermarked text is unexaminable, not human.

Synthetic test invoice · four findings converge on the total
A document keeps evidence of its own alteration — in the earlier versions saved inside the file, in a value drawn over another, in a figure set off its line's baseline, and in arithmetic that no longer adds up. Exiphore reads all of them, on PDFs, Word files and scanned or photographed pages, and draws each finding on the page.
The total reads 13,24,795.00; the figures above it add up to 3,24,795.00. The same figure is set in a different typeface 1.7 pt off its baseline, is drawn over the original value, and was changed in a later save.
What it cannot tell youDocuments are legitimately signed, stamped, completed and re-saved every day. A finding says the document changed, not that the change was dishonest — and authenticity is established only against the issuer's own record.
New · Document forensics
Bank statements, invoices, contracts — as PDFs, Word files, scans or phone photos. Exiphore recovers the value a figure replaced, the clause the page hides and the total that no longer adds up, and draws each finding on the page.
MERIDIAN LOGISTICS PRIVATE LIMITED
GSTIN 27AABCM4471K1Z5 · Bandra Kurla Complex, Mumbai 400051
TAX INVOICE INV/2026/0917
This is a computer-generated invoice and does not require a signature.
No examination of this file could have returned a higher confidence, whatever it found. A platform that reported 95% on this exhibit would be lying to you.
Your environment
Run examinations within your own environment. Choose a deployment that fits your agency’s operational requirements.
Deployment & security detailsOptional external narrative generation can be disabled for isolated operation.
Exiphore is a forensic deepfake detection and synthetic media forensics platform. It examines video, images, audio, documents and text for AI generation and manipulation, then produces a signed examination report with chain of custody. It runs on your own servers, including air-gapped networks. Exiphore is an AI Anytime product, built in Hyderabad, India.
For casework, the best deepfake detector is the one whose findings survive challenge: it states its method and limits, keeps chain of custody, reports a confidence interval rather than a bare score, allows an inconclusive result and runs where the evidence is allowed to go. Exiphore is built to those criteria. Test any tool on exhibits from your own cases before relying on it.
Twenty measurements across eight independent families of evidence: physiological signals such as the pulse recovered from four regions of a face, frequency and generator fingerprints, temporal and lip-sync coherence, compression artefacts, metadata and content provenance. Most are signal measurements rather than machine learning, so each result can be explained to a reviewer.
Eight examination agents each measure the exhibit independently, with no shared state, so agreement between them means something. An adjudication step pools correlated signals into one finding with an interval and a quality ceiling. A named human examiner then reviews and signs it. The agents measure; the examiner decides.
Yes. Exiphore examines face-swapped and lip-synced video, generated and edited images, cloned voices and synthetic speech, and altered PDFs and scanned documents. Model-generated text can be screened too, but text results are triage only and are labelled that way.
Yes. Exiphore recovers earlier versions saved inside a PDF, compares what the page shows with what the file holds, checks each line's typography, and reconciles totals, tax, balances and dates — on PDFs, Word files and scanned or photographed pages. Each finding is drawn on the page with the value it replaced.
Admissibility is for the court and counsel to decide. Exiphore is built to support it: the original is hashed (SHA-256) and sealed on intake, every action is recorded in a hash-chained custody log, each measurement states what it cannot establish, and a named examiner signs the finding. The hash values support certification of electronic records under Section 63 of the Bharatiya Sakshya Adhiniyam, 2023.
Yes. Exiphore runs on-premises, in a private cloud or fully air-gapped, on standard servers without GPUs. No exhibit, model call or telemetry leaves the network. The only optional component that can call an external model writes the plain-language summary; it never sees a measurement and can be switched off.
Performance depends on media quality, compression and the manipulation involved, so a single accuracy figure is not a guarantee for your casework. Every finding reports its confidence interval and the ceiling the exhibit’s quality imposes. Our published validation states its sample and limits; the best test is a pilot on exhibits from your own closed cases.
Inconclusive is a reported outcome, not a failure. Exiphore shows which evidence conflicted or was too degraded, so an examiner can seek the original file, gather more context or record the limits of the examination instead of forcing a verdict.
Recorded meetings, call recordings and video-KYC captures can be examined as exhibits today, for executive impersonation, payment fraud and onboarding abuse. A direct Microsoft Teams connector is scoped work, described in our meeting-recordings case study.
Take the first step toward exhibits you can stand behind. Book a guided demonstration with your use case in mind.
Figures as published by the named sources. Exiphore does not generate threat statistics of its own, and does not present other people's research as product evidence — our measured performance is published separately, with its method.