With the rapid increase in the number of Internet of Things (IoT) devices and the amount of data they generate, the traditional cloud-based approach is gradually unable to meet the actual needs of many scenarios. Distributed collaborative machine learning (DCML) paradigms such as Federated Learning (FL) and Split Learning (SL) provide possibilities for effective use of decentralized data in edge-based IoT. However, critical challenges in terms of data privacy, heterogeneity, and constrained resources remain to be handled. Despite extensive efforts, current solutions still cannot address the above challenges simultaneously. Therefore, studies in this emerging research field remain inadequate. In this paper, we propose a hybrid framework for combining FL with SL, named Privacy-Preserving Split Federated Learning (PPSFL). It facilitates privacy protection with an appropriate model decomposition strategy and mitigates the negative impact of data heterogeneity by incorporating private Group Normalization (GN) layers into the network. Extensive empirical results demonstrate that PPSFL attains better performance than other state-of-the-art distributed collaborative learning methods on different datasets. We also evaluate and compare the resistance of all baselines to reconstruction attacks with various image datasets. Results supported by comparative experiments indicate that our method can greatly prevent information leakage from raw data while maintaining classification performance. Additionally, the comparisons in terms of communication and computation overhead show that PPSFL is also competitive.

中文翻译：

---

PPSFL：基于边缘的物联网中异构数据的隐私保护分割联合学习

随着物联网设备数量及其产生的数据量的快速增加，传统的基于云的方式逐渐无法满足很多场景的实际需求。联邦学习 (FL) 和分割学习 (SL) 等分布式协作机器学习 (DCML) 范例为在基于边缘的物联网中有效使用分散数据提供了可能性。然而，数据隐私、异构性和资源有限等关键挑战仍有待解决。尽管付出了巨大的努力，当前的解决方案仍然无法同时解决上述挑战。因此，这一新兴研究领域的研究仍然不足。在本文中，我们提出了一种将 FL 与 SL 相结合的混合框架，称为隐私保护分割联邦学习 (PPSFL)。它通过适当的模型分解策略促进隐私保护，并通过将私有组标准化（GN）层合并到网络中来减轻数据异构性的负面影响。大量的实证结果表明，PPSFL 在不同数据集上比其他最先进的分布式协作学习方法获得了更好的性能。我们还评估和比较了所有基线对各种图像数据集的重建攻击的抵抗力。对比实验支持的结果表明，我们的方法可以在保持分类性能的同时极大地防止原始数据的信息泄漏。此外，通信和计算开销方面的比较表明 PPSFL 也具有竞争力。