Abstract
Face authentication has become one of the most widely used biometric technologies due to its high accuracy, convenience, and contactless operation. With rapid advancements in artificial intelligence and computer vision, face authentication systems have found extensive application in various critical fields, such as mobile device unlocking, online payment verification, access control systems, and surveillance systems. However, in recent years, numerous attack cases have revealed vulnerabilities in face authentication systems. Attackers may deceive these systems using photos, videos, or highly realistic 3D masks. Additionally, the widespread adoption of face authentication has raised privacy concerns. Unauthorized institutions collecting users' facial data may put this information at risk of being misused for malicious purposes, thus increasing the threat of privacy breaches and security risks.This dissertation addresses the limitations in security, practicality, usability, and privacy found in prior arts by proposing multiple security enhancement schemes tailored to different scenarios. These schemes can be flexibly applied based on scenario requirements or integrated into a multi-cue based system to defend against emerging attack threats. The main contributions of this dissertation are as follows:
Firstly, a security enhancement scheme leveraging gaze cues is proposed, tailored for camera-based face authentication. Exploiting the predictability and individual uniqueness of gaze patterns, this scheme provides high security guarantees without requiring additional hardware or imposing cognitive load on users. Specifically, this scheme displays random red dots (named gazecode) on the screen and captures predictable gaze responses and unique periocular deformations. By considering critical gaze information, this scheme proposes a gaze estimation model to perform accurate and fast gazecode verification. Additionally, a deformable periocular verification model is designed to extract stable and hard-to-forge individual features from periocular deformations to verify users' correctness. Evaluations on a 50-people dataset show that the system achieves detection rates of 95.72%, 95.59%, and 99.73% against image, video, and 3D mask attacks, respectively, demonstrating significant security enhancements.
Secondly, for mobile device-based face authentication, a security enhancement scheme leveraging gripping characteristics is proposed. This scheme leverages gripping characteristics as a second authentication factor to enhance the security of mobile face authentication while maintaining user convenience. During the authentication process, this scheme emits predesigned acoustic signals to capture the user's gripping characteristics on the device. Signal and feature extraction algorithms are then designed to retrieve structural transmission and hand echo signals, from which gripping features are derived. A deep learning model is proposed to verify the features and decide whether the current user is legitimate. To adapt to the changes in gripping patterns over time, this scheme proposes an incremental learning-based mechanism to ensure long-term effectiveness. Experimental evaluations on a 30-people dataset show that the system achieves a verification accuracy of 97.67% for legitimate users and an attack detection rate of 98.08% using the gripping gesture for face authentication.
Thirdly, for non-visual face authentication, an acoustic-based 3D dynamic face authentication scheme is introduced. This scheme presents a non-visual face authentication method that leverages acoustic signals to capture 3D dynamic facial features, addressing privacy concerns while enhancing security. During each authentication session, this scheme transmits predesigned acoustic signals to capture both the facial 3D structure and dynamic characteristics. Different signal processing and feature extraction algorithms are proposed to retrieve static and dynamic facial features from the received signals, respectively. A deep learning model combining Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks is proposed to integrate the static and dynamic features, performing secure and reliable 3D dynamic face authentication. Experiments show that the proposed method achieves an authentication accuracy of 94.45% on a dataset of 30 users, with an average attack detection error rate of 3.06%.
| Date of Award | 6 Jan 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Cong WANG (Supervisor) & Qian Wang (External Supervisor) |