How Can Crypto Exchanges and Digital-Asset Platforms Stop Deepfake KYC Fraud?

How Are Deepfakes Used Against Crypto Exchanges and Fintechs?

Deepfakes are used against crypto exchanges and fintechs in three ways: to open accounts under false or synthetic identities, to take over existing accounts, and to persuade real customers to send funds to scammers. Each one exploits a remote step where the platform trusts a face or a voice it cannot inspect in person.

Account opening is where regulators have the most data. The US Financial Crimes Enforcement Network said in alert FIN-2024-Alert004 on 13 November 2024 that criminals have used generative AI to alter or create ID photos and selfies, sometimes combined with stolen or invented personal data, to pass customer identification checks. Those accounts were then used to receive and launder proceeds of other frauds. FinCEN's red flags include a new account with rapid transactions and high payment volumes to digital asset exchanges, and withdrawals moved quickly to offshore exchanges where they are hard to reverse.

Account takeover follows the same pattern later in the customer's life. An attacker who has phished a password still needs to pass a selfie re-verification or a support call to reset two-factor authentication or add a new withdrawal address, and a face swap or cloned voice is aimed at exactly that step.

The third route bypasses the platform's controls entirely. On 14 October 2024 Hong Kong police said they had arrested 27 people linked to a syndicate that used AI-generated images and deepfake personas on social media to pitch cryptocurrency investments, with losses of about HK$360 million. The FBI's December 2024 public service announcement describes the same playbook and tells the public never to send cryptocurrency to someone they have not met in person.

Why Do Irreversible Transfers Change the Verification Design?

A card payment can be charged back and a bank wire can sometimes be recalled, but an on-chain transfer to an address the attacker controls is usually final. That shifts the cost of a missed deepfake from a dispute to a permanent loss, and it means the moment that matters most is often not sign-up but the first large withdrawal to a new address.

Many platforms verify a face carefully once and then rely on passwords and one-time codes for everything after. Attackers know this, so they aim synthetic media at the cheaper later steps: a support-led account recovery, a device change, a request to lift a withdrawal limit, or reactivation of a dormant account with a balance. Each of these should trigger a fresh capture and a fresh detection check, compared against the face enrolled at onboarding.

Identity and wallet risk also need to be kept apart. A genuine person can pass every biometric check while acting for someone else, funding the account with stolen money or moving assets through high-risk addresses. The verification design should record each answer separately rather than letting a good selfie score lower the wallet-risk rating.

How Can Digital-Asset Platforms Strengthen Remote Verification?

Platforms strengthen remote verification by checking media at every step where a face or voice releases funds, using capture controls that resist injected video, and sending unresolved cases to people trained on crypto fraud. The table below maps those steps.

Thresholds should differ by step. A borderline score at onboarding might justify a second selfie with a different pose, while the same score on a request to whitelist a new withdrawal address for a large balance should pause the request until a reviewer has looked at it. Record the score, the signals that drove it and the reviewer's decision, so the team can tune thresholds from real outcomes rather than guesses.

The capture itself matters as much as the model. Use the vendor's mobile SDK where possible, reject sessions that come from virtual cameras or emulators, and treat a customer who repeatedly reports camera problems or asks to finish verification on another channel as a risk signal, as FinCEN's alert recommends.

What Can a Deepfake Check Not Settle in Crypto KYC?

A deepfake check answers one question: does this face, video or voice show signs of AI generation or manipulation? It cannot show that the customer owns a wallet, explain where their funds came from, or clear an address that appears on a sanctions list. Those answers come from blockchain analytics, source-of-funds review and screening against lists such as those maintained by the US Office of Foreign Assets Control.

Virtual asset service providers also have obligations under the FATF's risk-based guidance, including customer due diligence, ongoing monitoring and the travel rule for originator and beneficiary information. A media check is one input to those processes. It should never be the reason an application skips them.

Detection also does little against the persona scams described above, because the deepfake is shown to the victim on a social platform, not submitted to the exchange. Platforms can still help with withdrawal warnings, cooling-off periods for first transfers to new addresses, and customer education, as the FBI recommends.

How Neural Defend Fits a Digital-Asset Platform

Neural Defend's NeuroVerify checks a face image, video or audio clip in one API call and returns a single risk score with signal-level evidence from six signals: visual safety screening, live presence verification, face manipulation detection, AI generative model analysis, AI pixel-level manipulation and AI content compliance. Your policy engine decides what each score means at each step.

Eligible image requests return in under 200 ms at p95 under defined test conditions, which keeps a step-up check on a withdrawal or recovery flow fast enough that honest customers do not notice it. Mobile SDKs for Android, iOS and Flutter handle capture, and backend SDKs for Python, TypeScript, Go and Java plus a REST API connect it to your onboarding and case systems. Because NeuroVerify sits next to your current provider rather than replacing it, you can follow the pattern in our guide to adding Neural Defend to an existing KYC stack.

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. The benchmark covers images only, and one author also co-wrote Neural Defend's ARCAS 1B paper, so test video and injection resistance on your own traffic. Platforms with data residency requirements can run NeuroVerify in a private cloud, on-premises or air-gapped, and Neural Defend holds ISO/IEC 27001, ISO/IEC 27701, ISO/IEC 42001 and SOC 2 Type II. NeuroVerify does not analyse ID documents.