Face Recognition on Embedded Systems in Compliance with the European General Data Protection Regulation

Abstract

The field of artificial intelligence (AI) is expanding rapidly and now encompasses many areas of daily life. Face recognition, in particular, has emerged as key technology in recent years. It is used in security systems, access controls and user authentication. The ability to detect and identify faces in images enables a wide range of applications, but it also raises significant data protection concerns, especially in the European Union. With the General Data Protection Regulation (GDPR) and the EU AI Act, the EU has established two major regulatory frameworks that govern the use of personal data, including biometric data information. These regulations impose strict requirements on the deployment of AI systems, including the need for clearly defined purposes and technical safeguards. At first glance, face recognition and data privacy may appear incompatible. These systems inherently require the processing of sensitive data. The aim of this thesis was therefore to investigate whether it is possible to develop a face recognition system that complies with the GDPR and AI Act while being deployable on resource-limited embedded platforms. The findings demonstrate that a privacy-compliant implementation is achievable and provides that the system architecture is built with data protection in mind, following the “privacy by design” principle. This includes, for example, the use of homomorphic encryption for pseudonymization and transparent data processing mechanisms. For the practical implementation, three different embedded systems were selected and evaluated. The system based on the i.MX8M Plus processor delivered the best performance in terms of inference speed. In contrast, the STM32MP2-based system achieved the highest accuracy with the lowest energy consumption, making it especially suitable for energy-efficient applications. Surprisingly, the NVIDIA Jetson Xavier NX system showed the weakest performance in this comparison, mainly due to the lack of hardware acceleration for TensorFlow Lite model execution.

Date of Award	2025
Original language	English
Supervisor	Franz Leopold Wiesinger (Supervisor)