Biometric Authentication Under Threat: Liveness Detection Hacking

Overview: Researchers from Tencent Security Keen Lab present technical methodologies for circumventing biometric authentication systems deployed across mobile devices and financial services. The presentation examines how contemporary biometric pipelines perform liveness and attention detection to prevent identity spoofing. The authors detail a hardware-level signal injection framework that directly inputs synthetic video and audio feeds into device sensor buses, bypassing optical and acoustic defense mechanisms.

Key Topics: - Biometric verification pipeline: Sensor collection, pre-processing, liveness detection, and feature matching - Optical, textural, and interactive liveness checks versus 3D physical spoofing - Limitations of presentation attacks: Color space loss, focus blur, and acoustic reverberation - Hardware-level video (HDMI to MIPI CSI) and audio signal injection architectures

Technical Details: Biometric authentication systems on modern smartphones and real-name identification apps process user verification through sequential stages: sensor data collection, digital pre-processing, liveness detection, and final biometric feature matching against stored templates. Liveness detection mechanisms verify that the biometric subject is a living person physically present before the sensor. Common defenses include human skin texture analysis, optical flow tracking, material frequency response, hardware infrared depth mapping, and interactive challenges requiring the user to blink, open their mouth, nod, or shake their head. Additional layers include attention detection, which verifies whether the user is actively gazing directly at the front-facing camera.

Prior research on bypassing facial recognition relied primarily on physical presentation attacks, such as displaying photos, replaying video recordings, or constructing complex 3D facial masks. However, these methods suffer from severe physical degradation: camera re-recording causes color space loss (in RGB and HSV spectrums), display moiré patterns, focus blur, and unnatural optical flow, while audio replays introduce room reverberation and inconsistent background acoustic noise. Furthermore, 3D facial molds are expensive and exceptionally difficult to reproduce reliably.

To eliminate optical and acoustic degradation, the researchers developed a transparent, hardware-level signal injection platform. Instead of presenting artifacts to external sensors, the attack intercepts the device's internal sensor buses. For video injection, an attacker PC generates synthetic video frames and routes the digital output through a hardware converter that transforms HDMI signals into native MIPI CSI (Camera Serial Interface) signals, connecting directly into the smartphone's camera connector socket. For audio verification, the PC routes synthesized voice responses directly into the microphone hardware pins. By monitoring interactive screen prompts over a secondary debug line, the attacker executes software on the PC to animate a single 2D static portrait in real time—generating head turns, nods, and blinks—and streams uncompressed native sensor data directly into the mobile processor with zero latency, completely bypassing liveness and attention checks.

Key Takeaways: Biometric security architectures cannot rely exclusively on optical and acoustic software checks if the underlying sensor hardware buses remain unauthenticated and physically accessible. Device manufacturers must establish end-to-end cryptographic mutual authentication and encrypted data transport between camera/microphone sensor modules and the system application processor. Multi-factor verification combining biometric input with hardware-backed cryptographic tokens should be mandated for high-value financial and real-name authentication transactions.

(AI summary — Gemini Flash)