You’ve probably seen these numbers separately: an eight-fold jump in synthetic identity fraud, and a deepfake attack every five minutes. Read alone, they look like headlines. Read together, they point to something structural: AI has turned identity fraud into an industrial process — the AI fraud front of the identity crisis.
The real question is whether the defences you already have can keep up. Let’s work through what’s changing, where the defences break, and what earns its place in your stack.
What is synthetic identity fraud and how is AI changing it?
Synthetic identity fraud is the fabrication of a person who doesn’t exist, stitched together from real stolen data, an ID number or address, and AI-generated names, headshots and details. The hybrid, sometimes called a Frankenstein identity, passes early KYC and credit checks. Unlike classic identity theft, there’s no human victim to complain, so losses surface months later.
AI has changed the economics. Generative models now automate all four steps of a scheme: creating the identity, forging documents like utility bills and licences, applying for accounts and credit, and passing liveness checks. AI fraud agents can even complete an onboarding flow, holding natural conversation and clearing verification without a human operator.
Fraud-as-a-Service then sells that capability as a subscription, turning a skilled attack into a volume business. LexisNexis reports an eight-fold, year-over-year increase, with synthetic identities behind more than one in ten frauds, and Deloitte predicts $23 billion in losses by 2030. It’s the front edge of the identity crisis.
What are deepfake identity attacks and how quickly are they growing?
Synthetic identities get a fake person in the door. Deepfakes widen the attack to the live human in front of the camera.
Deepfake identity attacks use AI-generated face and voice to impersonate a real person during remote onboarding, video KYC, and help-desk calls. The two main tactics are face swapping and voice cloning: one superimposes a victim’s face onto an attacker’s head, the other reproduces a voice from seconds of audio.
They beat live checks through injection: attackers replace the camera path with a virtual feed, so the verification system sees a synthetic stream rather than the real webcam. That’s how deepfakes beat live KYC, not just static images.
Deepfake fraud attempts surged 2,137% over three years, and Entrust recorded one attack every five minutes in 2024, a 31x year-on-year jump. iProov saw injection attacks rise 783% in a year, and Pindrop measured a 680% rise in voice-deepfake activity. Fraud-as-a-Service attack kits now sell for as little as US$5. It’s the same identity crisis, moving faster.
Why do traditional rule-based fraud detection systems fail against AI-driven fraud?
Both fronts now hit the same weakness: rule-based detection.
Rule-based detection runs on fixed thresholds and “if A, then B” logic, so it only catches patterns someone has already written down. That failure is structural. Generative AI produces fraud that looks like nothing the system has seen before.
Fraudsters engineer attacks to sit below the thresholds, spacing transactions and varying amounts so the pattern blends into normal behaviour. This low-and-slow style rarely trips a velocity rule. The signal fragments further, because transaction monitoring, identity verification and device analytics live in separate systems, so a coordinated campaign is never seen as one pattern.
The hidden cost is false positives: manual investigation can consume up to 22 hours per alert, producing alert fatigue and burnout. The counterweight is anomaly-based detection that learns normal behaviour, correlates context across silos, and flags deviations instead of waiting for a known signature, the adaptive layer in the identity stack.
How should I assess my organisation’s exposure to synthetic identity fraud?
That’s the shape of the threat. The rest is about the response, starting with your own exposure.
Before you buy anything, work out where the losses would land, because a synthetic identity that passes once tends to pass everywhere. Exposure concentrates on four entry surfaces: onboarding, where a fabricated identity gets in; KYC thresholds, where it clears screening; account recovery, where a cloned voice can reset credentials, the same gap token theft already exploits; and high-value actions, where the fraudster cashes out.
Quantify it strategically rather than as a compliance tick: which flows create accounts, credit or access, and what does one bust-out cost your business? For FinTech and HealthTech, AML, KYC and data-protection obligations turn a synthetic breach into a board-level problem. The Federal Reserve’s Synthetic Identity Fraud Mitigation Toolkit is the vendor-neutral reference.
What should I look for when evaluating an identity verification vendor against AI-generated fraud?
Once you know where the weak verification moment is, the vendor question becomes sharper.
Ask instead: “Can you pass our AI-generated samples within a week, and how fast do your models update?” The bar is a 7-Day Benchmark: AI-generated identities, forged documents and deepfake samples run through the vendor, scored against the attacks you craft rather than the samples it brings to the call.
The signals that matter are liveness detection, synthetic identity detection, document forensics, device and behavioural intelligence, configurable risk scoring, and a stated model-update cadence. Also check the layers are independent. If two detectors look for the same thing with different methods, that’s duplication, and it adds no defence-in-depth.
Vendors like Regula, Entrust, Sumsub, Jumio and Veriff are a testing set, not a recommendation. There’s no universal winner: adversarial robustness and update cadence matter more than benchmark accuracy, and these choices belong inside the defences across the identity stack.
How does traditional identity verification differ from liveness and biometric checks against deepfakes?
Those signals only matter once you understand what the older checks miss. Traditional verification is a document check plus a selfie match. It confirms a match between document and selfie, but says nothing about whether the person is live and present, so synthetic documents and deepfake injection walk through. Liveness and biometric checks instead ask “is this a live, real human, right now?”, using presentation attack detection, challenge-response prompts, and texture and temporal analysis.
Active liveness (blink, turn, follow a prompt) catches basic photo and replay attacks but is bypassed by real-time face swaps. Passive liveness is frictionless but locked in an arms race with generative models. The tradeoff: completion rates jumped from around 60% to over 95% after one organisation switched to passive liveness. Over 70% of advanced fraud attempts need multiple layers to stop.
But liveness is not a silver bullet. It has to sit inside a multi-layered stack of document, device, behavioural, liveness and risk-scoring signals. For high-value events you can add a non-camera modality like palm vein, which reads the vein pattern beneath the skin and has no public dataset a generator could train on.
That stack has to reach beyond authentication, since a single factor is easily defeated, and beyond the identity governance vacuum attackers exploit. The full picture is the post-password identity stack.
Conclusion
AI has industrialised fraud and outpaced single-point defences. A victimless fraud type and a real-time impersonation front have moved faster than static, rule-based detection.
The fix is architectural: those layers working together, chosen through sceptical, time-boxed testing. Assess exposure first, then run your own 7-Day Benchmark, and judge every vendor on the attacks it passes and how fast its models update when the attacks change. That’s one fault line in the post-password identity picture.
Frequently Asked Questions
Who actually loses money when a synthetic identity is created?
Usually the lender, the platform, or the organisation that onboarded the fake identity. Because no real person’s identity is stolen, there is no consumer to flag the account, so the fraud runs quietly until the account busts out and defaults. The loss lands on the business that extended credit or access, often months after the account looked legitimate.
How much does it cost to create a convincing deepfake today?
Less than most people assume. Generative AI and Fraud-as-a-Service tooling have collapsed the cost, time, and skill barrier, turning a face swap or cloned voice from a costly speciality into a low-cost commodity that can be bundled into a subscription attack kit. That cost collapse, not better fakes alone, is what has driven the surge to one deepfake attack every five minutes.
Is my own face or voice at risk of being stolen and cloned?
Yes, and the risk starts with material you have already shared. A short voice sample from a phone call or a face photo from social media can seed a clone. You cannot fully prevent sampling, but you can limit what is public, tighten privacy settings, and treat unsolicited requests that use a familiar voice or face with suspicion.
Can AI be used to catch the fraud that AI creates?
Yes. The same pattern-learning capability that generates novel fraud can also flag it. Anomaly and AI-powered detection systems learn normal behaviour and flag deviations instead of waiting for a known signature. The catch is that this is an arms race: detection models must be retrained on fresh attacks, which is why model-update cadence matters so much in vendor evaluation.
Do these attacks only affect banks and financial institutions?
No. Any organisation that remotely onboards customers or grants access to accounts, credit, or data is exposed. FinTech and HealthTech carry obvious AML, KYC, and data-protection obligations, but insurers, telcos, government services, and gig platforms all issue value through a remote identity decision. Wherever a weak verification moment exists, synthetic identities will find it.
What is a “bust out” in synthetic identity fraud?
It is the moment a synthetic identity stops behaving like a patient customer and cashes in. The fraudster builds the account’s history and credit limit over time, then maxes out the account and disappears. Because no real person exists to complain, the bust out is often the first signal the organisation gets, which is why these losses surface so late.
Is multi-factor authentication enough to stop AI-driven identity fraud?
No. MFA proves possession of a credential or device, not that the human behind the session is legitimate. A synthetic identity can pass MFA during onboarding, and deepfake or token-theft techniques can clear step-up prompts later. MFA is one useful layer, but it cannot replace liveness, document, device, behavioural, and risk-scoring signals working together.
What should I do to protect my own identity from AI fraud?
Limit what is publicly available, because face and voice samples are the raw material for deepfakes. Tighten social media privacy settings, avoid posting high-quality voice recordings, and be suspicious of unexpected calls that ask for account changes. If an organisation offers stronger verification options such as biometric or hardware-based login, enable them.
What happens if a synthetic identity passes our checks and we only find out later?
You absorb the loss and, depending on the sector, the compliance exposure. By the time a synthetic account busts out, credit has already been extended and the fraudster has gone. The practical response is to treat late discovery as a detection signal, trace where the weak verification moment was, and feed that pattern back into your layered controls and vendor benchmark.
Do smaller businesses need the same layered defences as large banks?
Proportionally, yes, because fraud now scales downward. Fraud-as-a-Service makes sophisticated attacks affordable against smaller targets that large institutions are better resourced to resist. The stack can be lighter, but the logic is identical: close the weak verification moment, layer document, liveness, device, and behavioural signals, and test vendors with your own samples.
