How to Stop Deepfake Fraud: Controls That Work

How do you stop deepfake fraud?

You stop deepfake fraud by making sure no single piece of media can authorise money, access or identity on its own. That means four layers working together: verification procedures that do not trust the channel a request arrives on, detection at every point where media enters a decision, identity checks hardened against injected and replayed video, and a response plan that can freeze the outcome while a person checks.

The order matters. Procedures are cheap and catch the attacks detection might miss. Detection scales those procedures to volumes people cannot handle. Hardened identity checks close the technical route attackers use to feed synthetic video into an onboarding flow. Response limits the damage when something gets through, which it eventually will.

Where deepfake fraud hits organisations

Deepfake fraud concentrates where a face or voice stands in for a person: onboarding, account recovery, payment instructions, customer calls and hiring. FinCEN's alert of 13 November 2024 reports a rise since 2023 in suspicious activity reports describing deepfake media, often generated or altered identity documents used to open accounts that then receive and launder proceeds of other fraud. The same alert cites the widely reported February 2024 case in Hong Kong in which a finance worker paid out $25 million after a video call with a deepfaked chief financial officer.

Consumers are targeted in parallel. The FTC warned in March 2023 that scammers were using cloned voices in family emergency schemes, and the FBI's December 2024 public service announcement describes cloned voices used to reach bank accounts, fake video calls with supposed executives or police, and generated identity documents.

Controls that work even when the fake is convincing

Several controls stop deepfake fraud without anyone needing to recognise a fake. Call back on a number already on file before acting on any payment or credential change. Require a second approver for transfers above a set value, and put a short hold on first payments to new beneficiaries. The FBI recommends families agree a secret word or phrase; companies can use the same idea with pre-agreed verification questions for high-value instructions.

For identity flows, FinCEN points to multifactor authentication, including phishing-resistant MFA, and live verification checks as practices that reduce exposure. It also notes the tells of an attacker struggling with those checks: repeated claims of technical glitches, requests to switch channel mid-check, and third-party webcam plugins that can display pre-generated video. Those behaviours are worth scoring alongside the media itself.

Training still helps, but aim it at procedure rather than spotting. In a 2022 PNAS study, people identified real versus StyleGAN2-generated faces with 48.2% accuracy, close to chance, and only reached 59.0% after training with feedback. Teach staff to follow the callback rule, not to trust their eyes.

Where deepfake detection fits in the defence

Detection belongs at each point where media enters a decision: the selfie and liveness video at onboarding, the face or voice used for recovery, the call audio in a contact centre, and the files submitted as evidence in claims or disputes. FinCEN lists a customer photo or video flagged by commercial or open-source deepfake detection software as one of its nine red flags, which shows regulators expect the signal to exist.

A useful detector returns more than a yes or no. It should say which signal fired, such as face manipulation, a generative-model fingerprint or a liveness failure, so a reviewer can act quickly and a policy can treat a likely injected stream differently from a heavily compressed genuine selfie. It should also answer fast enough to sit inside the flow; a check that takes many seconds gets bypassed or moved after the decision.

What detection cannot stop on its own

Detection cannot stop an employee who follows a convincing instruction without checking, a mule who opens an account with their own genuine face, or social engineering that never uses synthetic media at all. It also produces errors in both directions, so every flag needs a defined route and every pass is a probability, not a guarantee. Treat the score as evidence for a policy decision and keep the procedural controls above in place.

How to tell whether your defences are working

Track a small set of numbers per channel: attack attempts flagged, confirmed fraud that got through, false positive rate on genuine customers, time from flag to decision, and the share of high-value payments verified out of band. Review confirmed misses in detail; each one shows which layer failed. Re-test detectors when attackers change tools or when you add a channel, and compare against the baseline you recorded before deployment.

How Neural Defend helps stop deepfake fraud

Neural Defend covers the three places deepfake fraud is usually decided. NeuroVerify screens face images, video and audio at onboarding and recovery in one API call across six signals, including Live Presence Verification, Face Manipulation Detection and AI Generative Model Analysis, and returns one risk score with signal-level evidence in under 200 ms at p95 on eligible image requests under defined test conditions. NeuroCall, currently in Beta, focuses on voice deepfakes in calls. NeuroDetect gives fraud investigators a case workspace with heatmaps, frame and audio-segment views and a downloadable forensic report for disputes, claims and law enforcement referrals.

On the public VendorBench-100 benchmark of 100 adversarial still images, Neural Defend ranked first of 36 systems with an MCC of 0.876, against 0.646 for the next vendor; one author also co-authored Neural Defend's ARCAS 1B paper, and the results are public for anyone to rerun. Deployment runs as a managed cloud API, private cloud or VPC, secure container, on-premises or air-gapped. Your policy decides what happens with each score.