Securing Face Liveness Detection Using Unforgeable Lip Motion Patterns

Man Zhou,Qian Wang,Qi Li,Peipei Jiang,Jingxiao Yang,Chao Shen,Cong Wang,Shouhong Ding

DOI: https://doi.org/10.48550/arXiv.2106.08013

2021-06-15

Abstract:Face authentication usually utilizes deep learning models to verify users with high recognition accuracy. However, face authentication systems are vulnerable to various attacks that cheat the models by manipulating the digital counterparts of human faces. So far, lots of liveness detection schemes have been developed to prevent such attacks. Unfortunately, the attacker can still bypass these schemes by constructing wide-ranging sophisticated attacks. We study the security of existing face authentication services (e.g., Microsoft, Amazon, and Face++) and typical liveness detection approaches. Particularly, we develop a new type of attack, i.e., the low-cost 3D projection attack that projects manipulated face videos on a 3D face model, which can easily evade these face authentication services and liveness detection approaches. To this end, we propose FaceLip, a novel liveness detection scheme for face authentication, which utilizes unforgeable lip motion patterns built upon well-designed acoustic signals to enable a strong security guarantee. The unique lip motion patterns for each user are unforgeable because FaceLip verifies the patterns by capturing and analyzing the acoustic signals that are dynamically generated according to random challenges, which ensures that our signals for liveness detection cannot be manipulated. Specially, we develop robust algorithms for FaceLip to eliminate the impact of noisy signals in the environment and thus can accurately infer the lip motions at larger distances. We prototype FaceLip on off-the-shelf smartphones and conduct extensive experiments under different settings. Our evaluation with 44 participants validates the effectiveness and robustness of FaceLip.

Cryptography and Security

What problem does this paper attempt to address?

This paper aims to solve a series of security problems in facial authentication systems, especially the problem of insufficient effective defense measures against various spoofing attacks. Specifically, the paper first evaluates the security of existing mainstream commercial face recognition services (such as Microsoft, Amazon, Face++ etc.) and typical liveness detection methods, and finds that these systems and services are vulnerable to a new type of 3D projection attack. This attack can dynamically reconstruct the user's facial information by using low - cost devices (such as laptops, mini - projectors and general - purpose 3D silicone masks), and can easily bypass the existing facial authentication services and liveness detection methods.

To address the above - mentioned security challenges, the paper proposes a new liveness detection scheme named FaceLip. FaceLip constructs an unforgeable liveness detection pattern by using the unique lip movement patterns of each individual. Specifically, FaceLip is converted into an active sonar through the acoustic system of the mobile phone, and captures and analyzes the subtle movements of the lips according to the acoustic signals generated by random challenges, so as to achieve liveness detection with strong security guarantees. This method can not only resist the existing complex 3D dynamic attacks, but also accurately capture the unique lip movements of users in long - distance scenarios, and can maintain high accuracy even in the case of large environmental noise.

The design of FaceLip mainly includes three modules:
1. **Audio Capturing**: Play imperceptible acoustic signals through the mobile phone, and use the microphone to record the reflected signals to capture lip movements.
2. **Motion Verification**: Detect whether the recorded signals contain "physical" lip movements, and ensure that the signals are not relayed or tampered with. By converting the acoustic system of the mobile phone into an active sonar, calculate the fine estimation of the subtle lip movements, so as to achieve millimeter - level measurement accuracy.
3. **Consistency Verification**: Extract the unique lip movement patterns of each individual, and verify whether the input liveness detection is consistent with the real user through an efficient binary classifier, ensuring that the "physical" information constructed by complex attacks cannot bypass the system.

In conclusion, by proposing FaceLip, the paper aims to provide a facial liveness detection scheme that is still robust and secure in a wide range of attack environments (including 3D projection attacks).