What breaks when banks rely on face matching without liveness detection?

What face matching can do, and what it cannot prove

Face matching is a biometric comparison control. It can tell you whether the presented face resembles the enrolled face closely enough to pass a threshold, but it does not inherently prove that the capture is happening in front of a live person or through a trusted sensor path. That distinction matters because matching and capture assurance solve different problems.

In banking onboarding and payment approval, the control is often used as a convenience layer for remote identity proofing or step-up verification. The weakness appears when the system accepts a convincing image or video without verifying the source of the signal. A strong match can still be entirely synthetic if the input channel is not protected against replay, injection, or manipulation.

When teams evaluate this control properly, they should ask whether they are validating identity resemblance only, or also validating the integrity of the capture event. A face can be real, but the session can still be fraudulent if the image was relayed from another device, a virtual camera, or a deepfake pipeline. The missing control is usually not better matching, but stronger presentation-attack resistance.

Why banks become exposed to replay and injection attacks

The operational failure is simple: if the app trusts whatever arrives at the camera boundary, an attacker can feed it a replayed video, a generated face stream, or a manipulated capture that preserves enough biometric features to satisfy the matcher. That creates a false positive path into onboarding, account recovery, or payment authorisation.

This is why liveness detection is not a cosmetic enhancement. It adds evidence that the subject is physically present and that the signal is harder to spoof. Without it, the bank may have biometric similarity, but not capture integrity. Identity Proofing and KYC Guide and Biometric Authentication and Verification Guide both reflect this split between verification and anti-spoofing.

That exposure becomes more serious when the bank uses face matching as a high-trust signal with weak fallback review. If a successful biometric match is treated as equivalent to human presence and session legitimacy, the control can be bypassed without stealing a password or token. The attacker is then attacking the trust model, not just the model accuracy.

What changes in onboarding, approvals, and fraud decisions

In practice, the break is not limited to one biometric screen. It can affect customer onboarding, step-up verification, payee changes, card activation, and transaction approval flows that rely on the same capture assumption. Once the trust boundary is crossed, downstream business logic may accept the session as authenticated enough to proceed.

That is why the right question is not “did the face match?” but “was the capture channel trustworthy enough for the decision being made?” A low-risk self-service journey may tolerate weaker assurance, while an account-opening or payment-approval workflow often cannot. The same biometric result can therefore mean different things depending on where it sits in the process.

NIST SP 800-63 Digital Identity Guidelines are useful here because they separate identity proofing, authentication assurance, and verifier expectations. For attack-side context, MITRE D3FEND and MITRE ATT&CK Enterprise Matrix help teams reason about spoofing, credentialless entry, and post-compromise abuse patterns.

Risk and Threat Considerations

Without liveness detection, face matching becomes vulnerable to presentation attacks, including replayed video, virtual camera injection, and synthetic media. The main risk is not just a failed login, it is a false assurance event that lets fraud proceed through a control the business believes is strong.

Failure mechanism: the system accepts a visually plausible face as sufficient evidence of presence, while the attacker controls the input stream or capture path. That removes trustworthy proof that the biometric sample came from a live, untampered interaction.

Impact: attackers can open accounts, pass step-up checks, or approve transactions while appearing to satisfy biometric policy. At scale, this can undermine fraud controls, increase chargebacks and manual review load, and erode confidence in remote onboarding flows.

Practitioner Guidance

What to prioritise: Treat liveness and capture-path integrity as part of the control, not as optional hardening. If the decision has fraud, funds-movement, or account-opening consequences, the bank should require stronger anti-spoofing evidence than a face match alone.

What to verify: Confirm whether the vendor or in-house stack tests for presentation attacks, replay, virtual camera use, and synthetic media, and whether the result is enforced before any trust decision is made. A match score by itself is not an assurance statement.

Common mistake: teams often tune the threshold, then assume the biometric is “good enough.” Threshold tuning does not fix input-channel trust. If the capture source is untrusted, higher similarity can simply make the fraud look more convincing.

Practitioner takeaway: The real control objective is not face recognition, it is trustworthy evidence that the face sample was captured live and through a path the attacker could not manipulate.

Related resources from NHI Mgmt Group

• What breaks when face-based authentication is deployed without liveness detection or device controls?
• What breaks when selfie-to-ID verification is used without liveness detection?
• What breaks when organisations rely on IAM without identity threat detection?
• What breaks when organisations rely on detection without enforcement?