Mobile Device Background Sensors: Authentication vs Privacy

The increasing number of mobile devices in recent years has caused the collection of a large amount of personal information that needs to be protected. To this aim, behavioural biometrics has become very popular. But, what is the discriminative power of mobile behavioural biometrics in real scenarios? With the success of Deep Learning (DL), architectures based on Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), such as Long Short-Term Memory (LSTM), have shown improvements compared to traditional machine learning methods. However, these DL architectures still have limitations that need to be addressed. In response, new DL architectures like Transformers have emerged. The question is, can these new Transformers outperform previous biometric approaches? To answers to these questions, this thesis focuses on behavioural biometric authentication with data acquired from mobile background sensors (i.e., accelerometers and gyroscopes). In addition, to the best of our knowledge, this is the first thesis that explores and proposes novel behavioural biometric systems based on Transformers, achieving state-of-the-art results in gait, swipe, and keystroke biometrics. The adoption of biometrics requires a balance between security and privacy. Biometric modalities provide a unique and inherently personal approach for authentication. Nevertheless, biometrics also give rise to concerns regarding the invasion of personal privacy. According to the General Data Protection Regulation (GDPR) introduced by the European Union, personal data such as biometric data are sensitive and must be used and protected properly. This thesis analyses the impact of sensitive data in the performance of biometric systems and proposes a novel unsupervised privacy-preserving approach. The research conducted in this thesis makes significant contributions, including: i) a comprehensive review of the privacy vulnerabilities of mobile device sensors, covering metrics for quantifying privacy in relation to sensitive data, along with protection methods for safeguarding sensitive information; ii) an analysis of authentication systems for behavioural biometrics on mobile devices (i.e., gait, swipe, and keystroke), being the first thesis that explores the potential of Transformers for behavioural biometrics, introducing novel architectures that outperform the state of the art; and iii) a novel privacy-preserving approach for mobile biometric gait verification using unsupervised learning techniques, ensuring the protection of sensitive data during the verification process.