Abstract: Split Learning (SL) is an emerging privacy preserving technique that enables resource-constrained edge devices to participate in model training by partitioning models into client-side and server-side sub-models. While SL can reduce computational load on edge devices, it encounters significant challenges in heterogeneous environments with varying computing resources, communication capabilities, environmental conditions, and privacy requirements. Recent heterogeneous SL frameworks optimize split points for devices with varying resource constraints but often overlook personalized privacy requirements and local model customization under diverse environmental conditions. To address this, we propose P3SL, a Personalized Privacy-Preserving Split Learning framework for heterogeneous, resource-constrained edge systems. First, we design a personalized sequential split learning pipeline that allows each client to choose a customized split point and maintain local models tailored to its computational resources, environmental conditions, and privacy needs. Second, we formulate a bi-level optimization problem that empowers clients to determine their optimal split points without revealing private information (e.g., computational resources, environmental conditions, or privacy requirements) to the server. This approach balances energy consumption and privacy leakage while maintaining high accuracy. We have evaluated P3SL on seven edge devices—including four Jetson Nano P3450, two Raspberry Pi, and one laptop—using diverse models and datasets under varying environmental conditions. Results show that P3SL significantly reduces privacy leakage, lowers energy consumption by up to 59.12%, and consistently maintains high accuracy compared to state-of-the-art heterogeneous SL systems. [Download paper here](http://Fanwei100.github.io/files/2025001435.pdf) 