CISL: A Multiple Collaborative-Iterative-Distillation-Based Swarm Learning Framework for Internet of Vehicles

Abstract

As a data-free knowledge transfer paradigm, federated learning (FL) provides a novel solution for knowledge fusion in smart cities, especially in the field of Internet of Vehicles (IoV). However, the bandwidth bottleneck in the IoV limits the efficiency of federated collaboration, while trust issues associated with aggregation servers reduce users’ willingness to collaborate. To address these challenges, this article proposes a multiple collaborative iterative distillation-based swarm learning (CISL) framework for IoV. CISL leverages multiple collaborative iterative distillations to transform federated collaboration into serverless cross-device and cross-decentralized autonomous organization (DAO) knowledge transfer and fusion, enabling trustworthy swarm collaboration under bandwidth-constrained conditions. Moreover, it adaptively adjusts the inheritance and elimination of shared knowledge (SK) to enhance model adaptability and improve single-vehicle performance. Specifically, CISL proposes a collaborative iterative distillation mechanism that progressively integrates knowledge of other vehicles within the DAO, achieving cross-device SK fusion. Meanwhile, CISL introduces a multisage collaborative distillation mechanism, enabling each DAO to collaboratively distill and integrate SK from other DAOs, thereby expanding its knowledge domain. Additionally, CISL employs a dynamic balancing strategy to adaptively regulate the inheritance and elimination of SK, optimizing local models and enhancing their performance. Comprehensive experiments conducted on six benchmarks across two scenarios demonstrate that, compared to state-of-the-art methods, CISL exhibits superior adaptability and robustness across different datasets and task scenarios.