DEEPFAKE · AUDIO · VIDEO
How to detect audio or video deepfakes: an analysis guide
Deepfakes can imitate voices, replace faces or alter speech to make impersonation credible. At work, a fake executive request can pressure finance staff into bypassing normal approval under a pretext of urgency. The risk comes from psychological manipulation as well as media quality.
Learn how to use an audio and video deepfake detector as an examination aid, interpret technical clues and verify the request independently. A reassuring result is not payment authorisation.
1. Examine a suspicious audio recording
Interpret acoustic clues
An unusual voice, abrupt cuts or odd prosody may warrant checks, but synthetic speech may sound natural. Background noise, denoising, microphones and telephone transmission change the signal. Missing breaths or apparent regularity do not prove voice cloning.
Understand available spectral measurements
When the browser can decode the file, SecurCheck measures properties such as silence, clipping, amplitude variation and spectral concentration. A highly tonal signal may appear abnormal without being a cloned voice. These measurements are not a specialised model that authenticates speakers or certifies AI generation.
Connect compression with provenance
Container, codec, bitrate and available software markers help explain file processing. Heavy compression or cuts can also result from exporting, legitimate editing or a poor connection. Preserve the original file and avoid re-encoding it before examination.
2. Examine frames and audiovisual consistency
Look for anomalies without rushing to conclusions
Unstable edges, freezes or changing textures can warrant review. They can also result from video-call filters or compression. The module samples frames to measure some freezes, variations and block artefacts; it does not necessarily inspect every video frame.
Distinguish lip sync from technical delay
A mismatch between lips and voice can result from editing, latency or playback problems. Lip synchronisation requires specialised examination: SecurCheck does not produce an automated verdict on it. Apparently perfect synchronisation does not guarantee authenticity.
Do not treat visual details as definitive signatures
Interpret reflections, shadows, blinking and mouth movements in light of illumination, angle and media quality. Reflections in both eyes need not be identical in a genuine video. No single blink, texture or reflection reliably classifies a video as a deepfake.
3. Protect financial decisions against vishing
Confirm through an independent callback
Pause unusual, urgent or secret transfer requests. Contact the person through a number already known in your internal directory, not one supplied in the suspicious message. A familiar voice or face does not replace authorisation procedures.
Prioritise dual approval
Require two people to approve sensitive transactions and independently verify the beneficiary and bank details. A secret phrase can supplement controls, but may leak, be recorded or replayed: it is not impossible to bypass. Never use it as the sole proof of identity.
Preserve the source and report concerns
Preserve the native file, reception channel and useful timestamps. Metadata or a familiar codec does not prove sender identity, even on a business messaging service. Alert IT and finance when suspicious. If payment has already been made, promptly contact your bank and finance team.
Conclusion: interrupt urgency before acting
Technical analysis of suspicious media can help you pause, but no single clue or score authenticates an instruction. Preserve the file, verify the caller through another channel and apply payment controls even if the call seems entirely natural.
Deploy SecurCheck Business to support your finance and accounting teams