Government & Public Safety
Video, audio and images now arrive in case files having possibly been produced by software, and a result nobody can explain is worth very little the moment it is challenged. FaceOff examines what was submitted and returns graded findings with the reasoning kept alongside them, so an investigator, an expert witness and the other side can all examine the same thing
Overview
Examine video and audio submitted as evidence and produce findings with the reasoning behind them, so investigators can review and explain the results.
A case is decided on material that somebody produced: a recording of a call, footage from a camera, a photograph attached to a statement. For as long as anybody can remember, the argument about that material was an argument about what it showed. Increasingly it is an argument about whether it happened at all.
That is difficult for a process built on examination. A convincing forgery does not announce itself, and the people asked to judge it are lawyers, officers and judges rather than specialists in synthetic media. The reverse is just as damaging: a genuine recording can now be waved away by simply suggesting it was fabricated, and the suggestion costs nothing to make.
The checks below examine what was submitted and return a graded finding with the reasoning behind it, showing what was noticed and where. That is what lets a finding be reviewed by an expert instructed by the other side, explained to somebody who is not technical, and argued about properly rather than accepted or dismissed on trust.
What you deploy
These are the products investigators, prosecutors and legal teams use most often, covering the video, audio and images that arrive in a case file, including the recordings that are difficult to work with.
Key capabilities
What this sector asks for most often. Each one runs at the moment a decision is being made, and each returns a result a reviewer can explain to somebody else.
Recorded video is examined for face swaps and generated footage, and the finding points to the specific moments that need review. A long exhibit becomes a short list of passages a court or an expert can actually be taken through.
Recorded audio is examined for speech that was generated, cloned or altered rather than spoken, and the passages that raised concern are kept so they can be listened to rather than merely described in a report.
Nothing comes back as a bare yes or no. Results are graded and carry the findings that produced them, because a conclusion offered without its reasoning is the first thing to fall apart under cross examination.
A finding is written to be read by the person who has to rely on it rather than only by the person who produced it, so it can be put to a witness, explained in a hearing and reviewed months later by somebody who was not there at the time.
Real evidence is rarely clean. Recordings made in the open, over a telephone or with several people talking across one another can be separated by speaker and cleared of noise, so material that was going to be set aside becomes something a case can actually rely on.
Where a reference recording of a person exists, a questioned recording can be compared against it. That turns an assertion about who was speaking into something both sides are able to examine.
Where it is used
Every entry names a product on this site and describes it the way its own page does.
DeepVideoGuard checks recorded video for face swaps and generated footage, and points to the moments that need review.
DeepAudioGuard checks whether a recorded voice was spoken by a person or generated, cloned or altered.
VoiceExtract AI separates speakers and cleans up audio that is hard to work with, and compares a voice against a reference.
DeepImageGuard checks a still image for signs of editing or AI generation.
Common questions
A graded finding with the reasoning behind it, showing what was noticed and where in the recording. An investigator receives material that can be reviewed, explained and challenged rather than a verdict to be taken on trust.
No. Results are graded rather than binary, and every one carries the findings that produced it, so a reviewer can weigh it rather than accept it. A conclusion presented without its reasoning is the easiest thing in the world to attack.
VoiceExtract AI is built for difficult audio. It separates speakers and removes background noise and interference, so a recording that is hard to work with can become usable rather than being set aside as unhelpful.
Yes. VoiceExtract AI performs voiceprint matching and speaker verification, so a questioned recording can be compared against a reference recording of a person and that comparison examined by both sides.
Admissibility is decided by the court, under the rules of evidence that apply where the case is heard, and no supplier can promise it in advance. What can be done is to produce findings with the reasoning kept alongside them and a record of what was examined, which is what an expert on either side needs in order to review the work properly.
Yes, and it is meant to be examined. Each finding shows what was noticed and where, so an expert instructed by the other side can review the same material and reach an independent view rather than being asked to accept a score.
Yes. DeepImageGuard checks a still image in a case file for signs of editing or AI generation and highlights the areas that look manipulated.
DeepMeetGuard confirms that the person on a live video call is real while the call is happening, which applies to a remote hearing, a remote deposition or a client identification call just as it does to a business meeting.
Yes, and treating it otherwise would be the real mistake. Any method of detection can be mistaken in either direction, human judgement included, so a finding is evidence to be weighed rather than proof. That is precisely why the reasoning travels with it.
Bring your own footage. We will run it through the Adaptive Cognito Engine and walk you through the findings.
