Deepfake Statistics 2026: Fraud, Losses and Trends
Deepfake statistics for 2026 point the same way: attacks on identity checks are rising, single cases now cost hundreds of millions, and people cannot reliably spot fakes. According to a 2024 meta-analysis of 56 studies, human detection accuracy averages 55.54%, barely above a coin toss. Below, every figure also cites its primary source.
Deepfake Statistics 2026: Headline Numbers
Fraud, risk and compliance teams ask for these figures most often. Each one traces back to a regulator, law enforcement agency, analyst firm or peer-reviewed study. However, we exclude numbers that originate in identity verification vendors’ own reports.
| Statistic | Figure | Source and year |
|---|---|---|
| Growth in injection attacks on identity verification | +200% in 2023 | Gartner, 2024 |
| Enterprises expected to see identity verification as unreliable in isolation due to deepfakes | 30% by 2026 | Gartner, 2024 |
| US fraud losses enabled by generative AI | $12.3 billion in 2023, projected $40 billion by 2027 | Deloitte Center for Financial Services, 2024 |
| Executives reporting a deepfake incident targeting financial or accounting data | 25.9% | Deloitte poll, 2024 |
| Largest confirmed single deepfake video-call loss | About HK$200 million | Hong Kong government, 2024 |
| Average human accuracy at spotting deepfakes | 55.54% | Diel et al., meta-analysis, 2024 |
| Best model accuracy on unseen deepfakes in an open challenge | 65.18% | Meta Deepfake Detection Challenge, 2020 |
Deepfake Attacks on Identity Verification
Above all, the clearest signal comes from analyst and regulator data on how attackers target onboarding. According to Gartner (2024), presentation attacks remain the most common vector, but injection attacks rose 200% in 2023. According to the same Gartner release, some 30% of enterprises will no longer consider identity verification and authentication reliable in isolation by 2026, because of AI-generated deepfakes.
Meanwhile, regulators report the same direction. For instance, FinCEN, the US financial intelligence unit, said in a November 2024 alert that suspicious activity reports describing deepfake media had increased since 2023. In many cases, the reports involved altered or synthetic identity documents and selfies used to open accounts.
Police records also show how this looks in practice:
- Hong Kong, 2023. First, a loan fraud ring ran face swaps in at least 20 attempts, working from eight lost or stolen ID cards, according to Hong Kong Free Press coverage of the police case.
- Hong Kong, 2025. Then, according to the South China Morning Post, police said a ring used AI-altered ID photos to apply for 44 bank accounts, and 30 of those applications passed online checks.
In short, the pattern behind these numbers is consistent. Attackers moved from physical spoofs to injected and synthetic media, which liveness alone cannot judge. Our guide to liveness vs deepfake threats explains that gap.
Reported Financial Losses From Deepfake Fraud
In general, loss figures fall into two groups: single documented cases and modelled totals. Both are useful, but they answer different questions, so cite them carefully.
Documented Deepfake Fraud Cases
| Case | Year | Loss | Source |
|---|---|---|---|
| Multi-person deepfake video conference targeting a finance employee in Hong Kong | 2024 | About HK$200 million | Hong Kong government reply to the Legislative Council, 2024 |
| Romance and investment scam ring using deepfake video calls, Hong Kong | 2024 | HK$360 million, 27 arrested | Hong Kong police, as reported by The Record, 2024 |
| Face swap loan applications with stolen ID cards, Hong Kong | 2022 to 2023 | HK$200,000 | Hong Kong government, 2024 |
Modelled and Survey-Based Loss Figures
According to Deloitte’s Center for Financial Services (2024), fraud losses enabled by generative AI in the US could reach $40 billion by 2027, up from $12.3 billion in 2023. However, that projection covers all generative AI fraud, not deepfakes alone. Similarly, a November 2024 Federal Reserve toolkit put US synthetic identity fraud at about $35 billion for 2023, according to its own estimate, and it warned that generative AI makes such identities easier to build.
Survey data, meanwhile, adds the corporate view. In a 2024 Deloitte poll of more than 1,100 executives, according to the firm, 25.9% said their organisation had faced at least one deepfake incident targeting financial or accounting data in the previous year. Over half also expected such attacks to increase, 51.6% according to the same poll.
Deepfake Statistics for APAC and Southeast Asia
Regional data is patchier, because few agencies separate deepfakes from other scams. Still, police and government figures from across the region show deepfakes moving from novelty to routine tool.
| Market | Data point | Source and year |
|---|---|---|
| Malaysia | 454 fraud cases linked to deepfakes, with RM2.72 million in losses, January to August 2024 | Royal Malaysia Police CCID, reported by Bernama, 2024 |
| Indonesia | Deepfake videos of the president promoted a fake aid scam, so about 100 victims across 20 provinces paid fees | Indonesian National Police, reported by Tempo, 2025 |
| Vietnam | Online scam losses above VND 6,000 billion in the first 11 months of 2025, while deepfake video calls featured among the tactics | Ministry of Public Security, 2025 |
| Singapore | Central bank information paper on cyber risks associated with deepfakes | Monetary Authority of Singapore, 2025 |
| Hong Kong | Four separate police cases, for example in loans, account opening, romance scams and corporate payments | Hong Kong government and police, 2023 to 2025 |
However, two cautions apply. First, the Vietnam figure covers all online scams, so it is context rather than a deepfake total. Second, low reported numbers in a market often reflect how police classify cases, not low activity. For a regional view of the threat, see our overview of deepfake challenges in Asia.
Human vs Machine Deepfake Detection Performance
By now, the research on human detection is large enough to draw firm conclusions. In short, people do slightly better than chance at best, and they also overrate their own ability.
| Study | Participants or scope | Finding |
|---|---|---|
| Diel et al., meta-analysis, 2024 | 56 papers, 86,155 participants | Pooled accuracy 55.54%, while video reached 57.31%, images 53.16% and audio 62.08% |
| Nightingale and Farid, PNAS, 2022 | 315 participants | 48.2% accuracy classifying AI-generated faces |
| Köbis et al., iScience, 2021 | 210 participants | 57.6% accuracy on deepfake videos, although participants were overconfident |
| Groh et al., PNAS, 2022 | 15,016 participants | Crowds matched a leading detection model, but they made different mistakes |
| Meta Deepfake Detection Challenge, 2020 | 2,114 participants | Top model scored 65.18% on unseen videos, so generalisation remains hard |
Still, machines are not a silver bullet either. According to Meta’s 2020 challenge results, the best model dropped to 65.18% on a black-box set of videos it had never seen, well below its score on public data. In practice, that is why detection vendors retrain often and why detection works best as one layer among several. Our explainer on how deepfake detection works covers those layers.
Methodology and Source Notes for These Deepfake Statistics
We built this page for people who need to cite numbers, so the rules matter as much as the figures.
- Included: regulators, central banks, law enforcement, government records, analyst firms and peer-reviewed research.
- Excluded: figures that originate in identity verification or deepfake detection vendors’ own customer data, including when a third party repeats them.
- Dates: each figure carries its publication year, and we mark projections clearly.
- Scope: when a figure covers more than deepfakes, for example all online scams, the table says so.
- Refresh: we review this page every quarter, then replace figures older than 12 months when newer primary data exists.
So before you cite any number here, link to the original source, keep the year, and keep the scope. After all, a 2023 US projection quoted as a 2026 Asian fact is how weak statistics spread.
What Deepfake Statistics Mean for Identity Verification Teams
Three conclusions follow from the data. First, attackers have shifted toward injected and synthetic media. Second, people cannot be the main control, because their accuracy hovers near chance. Third, detection models help, but only when vendors retrain them and pair them with liveness and device checks.
That is also the design behind the VeriSecure SDK. It analyses one capture for liveness and deepfakes, while it screens the phone for emulators and injection tools. For flows that already collect selfies or videos, standalone deepfake detection can screen that media without changing the capture step.
Frequently Asked Questions About Deepfake Statistics
How common are deepfake attacks on identity verification?
Common and growing. According to Gartner (2024), injection attacks on identity verification rose 200% in 2023, and FinCEN reported more suspicious activity reports describing deepfake media from 2023 onward.
How much money do businesses lose to deepfake fraud?
Single cases have reached roughly HK$200 million, according to a Hong Kong government reply to legislators. According to Deloitte (2024), US fraud losses enabled by generative AI could also reach $40 billion by 2027, although that covers more than deepfakes.
Can people detect deepfakes?
Not reliably. According to a 2024 meta-analysis of 56 studies, average human accuracy was 55.54%, close to chance. People also tend to overestimate their own ability.
How accurate is deepfake detection software?
It varies, because results depend on the data. According to Meta’s 2020 challenge, the top model scored 65.18% on videos it had never seen. Accuracy holds up better when vendors retrain frequently and combine detection with liveness and device checks.
Which deepfake statistics are reliable to cite?
First, prefer figures from regulators, police, governments, analyst firms and peer-reviewed studies, and keep the year and scope. Also treat vendor-generated statistics with caution, because vendors rarely publish their methods.
Deepfake Statistics Show a Detection Problem, Not Just a Fraud Problem
Read together, then, the numbers tell a simple story. Deepfake fraud is no longer rare, single cases reach eight or nine figures, and the human eye is close to useless as a defence. Yet the weakest link is not awareness. Instead, it is any process that still relies on a person, or a single model, to judge whether a face is real.
So for identity teams, the practical takeaway is to treat deepfakes as a standing risk in onboarding and step-up flows, measure it with your own data, and layer controls accordingly. Meanwhile, we will keep this page updated each quarter as new primary data appears.
Want to see how your own onboarding data compares? Ask Verihubs for a deepfake risk review of your verification flow.
