This paper proposes P3SL, a personalized privacy-preserving segmentation learning framework for resource-constrained edge devices in heterogeneous environments. To address the inherent differences in resources, communication, environmental conditions, and privacy requirements in heterogeneous environments, which are inherent limitations of traditional segmentation learning (SL), we design a personalized sequential segmentation learning pipeline that allows each client to customize its own privacy level and local model. Furthermore, we use a bi-level optimization technique to enable clients to determine the optimal segmentation point without sharing their private information (computational resources, environmental conditions, and privacy requirements) with the server. We implement and evaluate P3SL using various model architectures and datasets in a test environment consisting of four Jetson Nano P3450s, two Raspberry Pis, and one laptop. We aim to achieve high model accuracy while maintaining a balance between energy consumption and privacy leakage risk.
