Anonymization of Images for Privacy Protection on Embedded Systems

As small, camera-equipped devices become increasingly common, they of- ten capture images containing faces and other personal details. This raises significant privacy concerns, especially under regulations such as the Euro- pean Union’s General Data Protection Regulation (GDPR), which require that such sensitive data be anonymized. However, implementing strong pri- vacy safeguards directly on these devices is challenging due to their limited computing power. In our research, we explore the trade-off between computational efficiency and privacy protection by implementing three lightweight anonymization techniques, namely masking, pixelation, and blurring, directly on a resource- constrained embedded platform, the ESP32-P4. Each method is evaluated for its execution time and its effectiveness at protecting identity, the lat- ter measured using cosine similarity scores derived from the DeepFace Python library. By analyzing the performance and privacy impact of these techniques, our work aims to uncover practical strategies for real-time, on- device anonymization. This enables privacy-preserving image capture at the edge, without the need for potentially insecure cloud processing. Our findings show that all three anonymization techniques can be exe- cuted in under 2 microseconds on the ESP32-P4, making them highly suitable for real-time processing. The full anonymization pipeline, including face ˜ detection, operates at 20 milliseconds per frame, enabling throughput up to 50 FPS. While detection accuracy reached 59.2% across a diverse dataset of 200 images, the main performance bottleneck lies in the face detection model, not in the anonymization methods themselves. These results confirm the viability of real-time, edge-based visual anonymization on constrained embedded systems