.png)
Somewhere this quarter, a bank's fraud team will catch a deepfake at onboarding and count it as a win. The same synthetic face, generated from the same breached photographs, will pass at a telco across town the following week, and at a mobile lender the week after that. Each institution will record its own result. None of them will see the attacker's ledger, which is the only place the campaign makes sense: one fake, produced once, sold or reused across every verification flow in the country until something stops it.
That ledger is worth taking seriously, because deepfake fraud at scale behaves like a business. It has production costs, distribution channels and a return on investment, and the documented returns are now substantial: upwards of $3.7 billion in global losses tied to deepfake fraud, with about 89% of it recorded in 2025 and the first half of 2026. Businesses respond to margins. A strategy that wants to beat this one should be asking what its margins depend on.
The margins of impersonation
On the cost side, the attacker's inputs keep getting cheaper. Generative tools have turned a convincing synthetic face or voice from a specialist effort into a commodity, and the raw material is often free: the photographs and personal details needed to fake a specific person sit in the same central biometric databases that identity systems maintain in order to verify people. Every breach of such a database is, in effect, an inventory delivery.
On the revenue side, the decisive variable is reuse. A fake face that works is not spent when it passes a check. In a system where every bank, telco and lender runs its own camera moment, the same fake can be presented at thousands of doors, and the attacker needs to find only the weakest of them. Worse, each successful pass generates authentic-looking records, so the fraudulent identity compounds in credibility as it moves through the economy (AlphaX, 2026). High reuse, near-zero marginal cost, appreciating asset: these are excellent unit economics, and they belong to the attacker.
Why better detection barely moves the price
The standard institutional response is to raise the cost of a successful fake through better detection: sharper liveness checks, smarter forensics, faster model updates. That work matters, and nothing in this argument removes the need for it at the moments where a camera genuinely must stand guard. But as a pricing intervention it is weak, because it acts on the one variable the attacker can most easily re-optimise. Generation tools improve on the same curve detection tools do, with more commercial energy behind them, and Entrust's finding that deepfakes now appear in roughly one in five biometric fraud attempts worldwide suggests the fakes are clearing real-world checks at scale already (Entrust, 2026, as cited in AlphaX, 2026).
Detection also does nothing about reuse. A fake caught at one institution remains available to every other, because the institutions are running separate checks against separate equipment with no shared memory of the attempt. The attacker's asset survives every local defeat. As long as that is true, the economics stay intact no matter how good any single camera becomes.
What actually collapses the market
The intervention that changes the business is architectural, and it works on the revenue line rather than the cost line. Verify a person once, at a single supervised and well-defended enrollment, anchor that event so it can never be quietly inserted or edited, and let every subsequent check verify a cryptographic proof of it. In that arrangement, a camera exists at exactly one point in the system, and it is the most defended point. Everywhere else, there is nothing for a fake face to be shown to.
Run the attacker's ledger again under those conditions. Production costs are unchanged; the tools remain cheap. But reuse has gone. A fake that beats the enrollment moment, which remains possible against a hardened, supervised, deduplicated process, yields a single anchored credential. That credential is traceable and revocable, so the asset can be seized. There is no central biometric pool to restock inventory from, and no way to backfill supporting records, because history on an anchored layer cannot be rewritten after the fact. The fake face still exists. It just no longer earns anything worth the effort, and campaigns that stop earning stop running.
This is also why the window matters more than any single procurement decision. Systems specified today will face materially worse attack conditions within 18 to 24 months. A government that spends those months buying incrementally better cameras is bidding up the attacker's costs by percentages while leaving their revenues untouched. The full structural argument, the data behind it and the deployments already running pieces of the model are laid out in our page on deepfake identity fraud, together with the market report it draws on.
Price the fraud out
Fraud prevention is usually framed as a contest of skill between attackers and defenders. The framing flatters both sides. What ends a fraud category, historically, is rarely a smarter guard; it is a change in the environment that leaves the crime unprofitable.
Deepfake identity fraud is unusually exposed to exactly that kind of change, because its entire return depends on one architectural accident: the thousands of repeated camera moments that identity systems never deliberately chose and do not need. Remove the accident, and the market for fake faces goes with it. That is a decision available to every national programme now, and it is a better use of the next two years than another round of the race.