A deepfake detector builtfor evidence, not for a score

There is no single best deepfake detector, because detectors differ in what they measure, what media they handle, and how they behave when the evidence is poor. A more useful question is whether a tool measures several independent families of evidence, reports a confidence interval rather than a single score, can return inconclusive, refuses to infer authenticity from artefacts it did not find, and produces a record an examiner can sign. Exiphore is built against all five, and publishes the conditions behind its own accuracy figure so it can be checked rather than taken on trust.

Free online deepfake detectors exist and are useful for casual curiosity, but they are not suitable for evidential work: they require uploading the material to a third party, return a single unexplained score, keep no chain of custody, and state no limitations. For anything that may end up in a case file, the output has to be defensible, and that means measurements, stated limits, custody and a named examiner. Exiphore is licensed to institutions and deployed on their own infrastructure.

Accuracy depends far more on the exhibit than on the tool. On a balanced 60-image sample of the public Hemg/deepfake-and-real-images corpus, measured end-to-end through its deployed pipeline, Exiphore flagged 86.7% of manipulated exhibits with a 0.0% false-alarm rate and returned undetected manipulations as inconclusive rather than clearing them. That is a small sample of still images from one corpus. Detection rates fall substantially on compressed, re-encoded and platform-redistributed media, and all learned detectors generalise poorly to generators absent from their training data.

Partly, and the honest answer is that redistribution destroys evidence. Re-encoding strips metadata, removes any provenance layer, and attenuates the frequency and compression traces a detector relies on. Exiphore still examines such exhibits — most casework material arrives this way — but it measures the damage and caps the confidence any result is allowed to carry, rather than reporting the same certainty it would on an original.

Synthetic media detector is the broader term for the same category. Deepfake usually implies a manipulated human face or voice, whereas synthetic media covers everything machine-generated: fully generated video with no real subject, cloned speech, AI-generated imagery, fabricated documents and synthetic identities. Exiphore examines all of these, and reports which family of evidence supported the finding in each case.

An automated detection result is not an expert opinion on its own, in any jurisdiction. What can be tendered is a forensic examination: the measurements taken, their stated limitations, a tamper-evident chain of custody linking the finding to a specific set of bytes, and a qualified examiner's recorded reasoning and signature. Exiphore is built to produce exactly that record, and is explicit in it about what the examination cannot establish.

In part, and the limit is worth stating plainly. Learned classifiers only recognise what resembles their training data, so a generator released after the model was trained may evade that signal entirely. This is why Exiphore does not rely on a learned classifier alone: physiological, frequency, temporal and provenance evidence do not depend on having seen a particular generator before, and a new method has to defeat all of them at once rather than just one.

No. Exiphore is installed on your own infrastructure and can run fully air-gapped. Evidence never leaves your estate, there is no hosted tier and no external service in the examination path. The only component that can call an external model is optional narrative generation, which never touches the verdict or any measurement and can be disabled entirely.

Visual inspection is unreliable and getting worse — the artefacts people are told to look for, such as unnatural blinking or blurred face edges, are exactly what each generation of tools fixes first. What is still worth doing before any technical examination is investigative rather than visual: preserve the original file rather than a re-shared copy, capture the source account and posting time, look for an earlier or higher-quality version of the same footage, and note whether the claim is corroborated anywhere else. Those steps hold up even when a detection result does not.