Facial biometrics in the AI era: Why banks must tighten device controls, lock down applications, and enforce multi-factor authentication

Starting from March 1, 2026, when Circular 77 issued by the State Bank of Vietnam officially mandates stricter control over device environments and Mobile Banking version management, users will increasingly notice tangible changes, such as banking applications being blocked on rooted or jailbroken devices, automatically terminating sessions when insecure environments are detected, or requiring mandatory upgrades to the latest application version. These developments indicate that the banking sector is proactively standardizing its security architecture ahead of the regulation’s effective date. Based on the current technical roadmap and implementation pace, there are strong indications that the 2026-2028 period will become a major security standardization cycle for digital banking applications in Vietnam – one in which biometrics no longer operate as a standalone authentication mechanism but instead function within a layered defense architecture and risk-based authentication framework.

Facial biometrics have become the default – but also the most attractive attack surface

Facial biometrics are rapidly becoming the default authentication method in digital banking due to their convenience, speed, and strong security characteristics. From a cybersecurity perspective, however, they also represent one of the most attractive attack surfaces. Once this authentication layer is bypassed, fraudsters may gain control of a transaction session and manipulate subsequent processes.

In practice, the primary risk does not reside solely in facial recognition algorithms themselves but in the deployment context and device environment in which they operate. Authentication experience studies in Vietnam have shown that 58.3% of users utilize biometric authentication, while one in three users expresses concerns about biometric theft or spoofing. The risk of identity fraud increases significantly in online environments through techniques such as deepfakes and voice cloning, particularly when biometrics are used as a standalone authentication mechanism.

1. Compromising biometric security often begins with the application runtime environment

A common misconception is that robust liveness detection alone can eliminate AI-generated attacks and deepfakes. In reality, many fraud scenarios take an indirect approach: instead of confronting facial recognition AI directly, attackers target the application or device environment to manipulate processing workflows.

As a result, the inevitable direction of risk management is clear: Mobile Banking applications must be capable of detecting insecure environments and automatically ceasing operation rather than attempting to function within compromised conditions.

In other words, if the runtime environment itself is under attacker control, the integrity of all biometric signals becomes questionable.

Common indicators of high-risk environments include:

• Applications running on emulators or virtual machines.
• Devices connected to debugging tools or technical intervention mechanisms, such as Android Debug Bridge (ADB).
• Applications subjected to hooking, API and function-call interception, or data flow monitoring.
• Repackaged or modified application binaries.
• Rooted or jailbroken devices.
• Devices with unlocked bootloaders.

This is why measures that restrict application usage on compromised devices are not intended to inconvenience customers; rather, they eliminate the fundamental conditions that enable attackers to interfere with authentication and transaction workflows.

2. Lessons from major banks: Restricting services on compromised devices

Several banks have already begun implementing these controls. For example, Vietcombank announced the discontinuation of VCB Digibank services on compromised iOS devices (including jailbroken or hooked devices) from November 2, 2024. The application is also configured to display warnings and automatically terminate when signs of tampering are detected.

Across the financial services and banking sector, technical requirements are increasingly being standardized in this direction. Beginning March 1, 2026, banking applications are expected to automatically terminate or suspend operations when insecure environments are detected, including the scenarios described above. In addition, institutions will be required to implement application version control, downgrade prevention mechanisms, and periodic security assessments at least once every three months to minimize the risk of exploitation through long-standing vulnerabilities.

The key objective is not to impose restrictions on users but to formalize banks’ responsibilities for securing the application runtime environment – the point at which actual security risks emerge.

3. AI and deepfakes make standardized liveness detection and layered security essential

AI-powered deepfakes are far more than realistic fake videos; they represent an ongoing arms race between synthetic content generation and fraud detection capabilities.

As a result, financial institutions must:

• Standardize biometric Presentation Attack Detection (PAD) requirements according to international standards, commonly referenced as ISO/IEC 30107 Level 2 or equivalent.
• Recognize a critical operational principle: PAD and liveness detection are only meaningful when both the application and runtime environment remain uncompromised. Once the device or application has been tampered with, liveness signals themselves may be manipulated or rendered unreliable.

4. Facial biometrics must be combined with risk-based multi-factor authentication (MFA)

Biometrics represent an inherence factor – something the user is. However, they should never serve as the sole key for authorizing high-risk activities.

Authentication experts consistently emphasize that biometric risk is closely tied to application context. In many implementations, biometrics function merely as a local unlocking mechanism that activates a separate authentication process behind the scenes.

In digital banking, effective MFA design follows a risk-based authentication model, escalating security requirements for sensitive events, including:

• New device registration and account login.
• Account recovery processes.
• Changes to identity information or authentication methods.
• High-value transactions or anomalous behavior, such as unfamiliar devices, unusual locations, or abnormal interaction patterns.

True MFA is not simply the addition of an OTP code. It requires combining at least two of the following three authentication factor categories:

• Possession: Smart OTP, soft tokens, registered trusted devices, or transaction signing mechanisms.
• Knowledge: PINs or passwords.
• Inherence: Facial recognition or fingerprint authentication supported by ISO-compliant PAD and liveness detection.

5. A defense-in-depth approach to biometric fraud prevention

If biometric fraud prevention in digital banking were distilled into a single security architecture, it would consist of five core control layers:

• Device and runtime environment controls: Detect root or jailbreak status, unlocked bootloaders, emulators, virtual machines, debugging tools, ADB connections, application hooking, and repackaging attempts.
• Application protection: Prevent application tampering, repackaging, runtime manipulation, and reverse engineering while ensuring data flow integrity.
• ISO-compliant PAD and liveness detection: Deploy ISO/IEC 30107-3 Level 2 or higher PAD capabilities recognized by FIDO Alliance standards to counter deepfake-based attacks.
• Risk-based MFA: Apply adaptive authentication for critical user actions, particularly device changes and account recovery scenarios.
• Version and vulnerability management: Enforce the latest secure application versions, implement downgrade prevention, and conduct security assessments at least once every three months.

Biometrics have not failed – Poor implementation creates the attack surface

Biometrics have not failed.

What creates vulnerabilities is the implementation of biometrics as an isolated security layer, disconnected from device trust controls, application integrity mechanisms, and version governance frameworks.

The 2026-2028 period is poised to become a defining era of digital banking application security standardization in Vietnam. During this phase, requirements surrounding device security, runtime protection, PAD standardization, and risk-based MFA will evolve from technical recommendations into mandatory industry standards.

In this environment, competitive advantage will no longer be determined by whether an institution offers biometric authentication. Instead, it will depend on the strength of the layered security architecture operating behind the biometric interface.