What deepfake detection is, and how to evaluate a detector

Deepfake detection is the automated analysis of audio, images, video or documents to decide whether they were generated or altered by AI. A detector returns a verdict and a confidence score, and the good ones also show where the manipulation is, so a person can act on the answer rather than take it on trust.

Detectors do not judge whether a story is plausible. They look at the signal. Synthetic speech leaves traces in its frequency distribution and vocal timing that survive compression and a bad phone line. A generated image carries statistical fingerprints of the model that made it, and an edited one has a seam where the pixels stop agreeing with each other. A face swap holds up frame by frame but slips between frames. A forged document changes in one region and stays untouched everywhere else. Each of those is a different problem with a different detector, which is why one model for everything tends to be mediocre at all of it.

People are poor at this task. The voice on the phone sounds like the finance director because it was built from the finance director's own recordings, and the face on the call moves the way faces move. The Hong Kong finance worker who transferred twenty five million dollars was on a video call with what looked like several colleagues, every one of them synthetic. Every incident on this site was carried out against someone who was paying attention. Detection exists because attention is not enough.

Raid AI runs a separate detector for each format behind one dashboard, one history and one API, so a verdict during a call and a verdict during the investigation afterwards come from the same place. Each product has its own page with the inputs it takes, what it catches and how it runs.

• —Voice and audio: cloned voices and synthetic speech in live calls and uploaded audio, scored on the acoustic signal rather than the words.
• —Image forensics: a generated or edited image told from a real one, with a heatmap of the altered region and the family of model that produced it.
• —Video forensics: frame level analysis for face swaps and temporal inconsistency in recorded video, run as a background job on long files.
• —Document forensics: forged and wholly generated documents in onboarding and claims, with the edited region boxed and the evidence kept beside the verdict.
• —Meeting protection: a verdict during the call on Teams, Zoom, Google Meet and Webex, while there is still time to act on it.

Eight questions that separate a demo from a deployment. They are written so that a reader can run the evaluation without us in the room, and they apply to us as much as to anyone else.

1. 01Test on your own media. Bring a recording, an image or a document you already know the answer to. A vendor demo on the vendor's samples proves the demo works.
2. 02Use the quality your traffic actually has. Phone audio at 8 kHz, a screenshot of a screenshot, a PDF scanned twice. Detectors that shine on studio files can fall apart on a real call.
3. 03Ask for the false positive rate. A detector that flags one honest caller in fifty will be switched off within a month. Accuracy alone hides this number; ask for both sides of it.
4. 04Try a language your customers speak. Most models are trained on English. If your calls are in Arabic, Hindi or Tagalog, test in Arabic, Hindi or Tagalog, and test dialects, not the broadcast standard.
5. 05Time the verdict against the decision. A verdict that arrives after the transfer clears is a report, not protection. For live calls the number that matters is time to verdict while the call is open.
6. 06Find out where the media goes. Which engines score it, in which country, for how long it is kept, and whether the whole stack can run inside your own network when the answer has to be nowhere.
7. 07Look for evidence, not only a score. A heatmap of where an image was edited, the region of a document that changed, the segment of a call that flipped. A bare percentage is hard to act on and harder to defend.
8. 08Check it reaches the place decisions are made. The IVR, the agent desktop, the meeting, the onboarding flow. A detector that lives in its own dashboard gets consulted after the fact, if at all.

The same product either way. The difference is where the media goes, and for some buyers that is the whole decision.

Cloud API

• —A key and a few lines of code; audio and image answer in one call, video is submitted and polled.
• —Official TypeScript and Python SDKs with retry, typed errors and generated types.
• —Retention you control: zero retention, or a window you set.
• —Media goes only to the engines your tenant is configured to use.

Your own infrastructure

• —The full detection stack inside your network, including air gapped environments with no outbound connectivity.
• —Media and results stored in your own database, so nothing is written to Raid AI storage.
• —The same models, the same dashboard, the same API surface as the cloud.
• —Built for regulators, sovereign wealth and defence, where the answer to where does it go has to be nowhere.

Accuracy gets a detector into the room. These are what keep it there.

• —99.7% detection accuracy across the conditions audio actually arrives in: phone calls, video conferences, recorded files and compressed streams.
• —A verdict in under 50 milliseconds, fast enough to flag a synthetic voice during a live call rather than after the transfer has cleared.
• —Over 120 languages, with accuracy on Arabic dialects that English first models do not have.
• —Retention you control: zero retention, or a window you set. Enterprise tenants can keep the record inside their own infrastructure.
• —Cloud or your own network: call the cloud API in a few lines, or run the full detection stack inside air gapped infrastructure.
• —Multimodal detection: voice, image, video and documents from one platform, covering voice cloning, face swaps, synthetic images, lip sync manipulation and forged paperwork.

Every organisation that acts on a voice, a face or a document is a target. Financial institutions face executive voice fraud and unauthorised wire transfers, where a single deepfake incident has cost a company over 35 million dollars. Government and defence need classified communications and diplomatic channels protected against voice impersonation, with deployment inside their own network. Media and journalism need to verify a recording before it is published and confirm that a statement came from the person it is attributed to. Enterprise communications on Teams, Zoom, Meet, Slack and Webex need to be checked for synthetic participants while the call is still running. The industries with a full breakdown of the attacks and the defences are on the use cases page.