[Paper Review] FedICT: Federated Multi-task Distillation for Multi-access Edge Computing

Zhiyuan Wu, Sheng Sun+5 more|arXiv (Cornell University)|Jan 1, 2023

Privacy-Preserving Technologies in Data5 citations

TL;DR

FedICT proposes a communication-efficient federated multi-task distillation framework for multi-access edge computing that enables personalized, heterogeneous model training without public datasets. By introducing Federated Prior Knowledge Distillation (FPKD) and Local Knowledge Adjustment (LKA), FedICT reduces communication overhead by less than 1.2% and training rounds by up to 75% compared to baselines, while mitigating client drift and improving performance under data and model heterogeneity.

ABSTRACT

The growing interest in intelligent services and privacy protection for mobile devices has given rise to the widespread application of federated learning in Multi-access Edge Computing (MEC). Diverse user behaviors call for personalized services with heterogeneous Machine Learning (ML) models on different devices. Federated Multi-task Learning (FMTL) is proposed to train related but personalized ML models for different devices, whereas previous works suffer from excessive communication overhead during training and neglect the model heterogeneity among devices in MEC. Introducing knowledge distillation into FMTL can simultaneously enable efficient communication and model heterogeneity among clients, whereas existing methods rely on a public dataset, which is impractical in reality. To tackle this dilemma, Federated MultI-task Distillation for Multi-access Edge CompuTing (FedICT) is proposed. FedICT direct local-global knowledge aloof during bi-directional distillation processes between clients and the server, aiming to enable multi-task clients while alleviating client drift derived from divergent optimization directions of client-side local models. Specifically, FedICT includes Federated Prior Knowledge Distillation (FPKD) and Local Knowledge Adjustment (LKA). FPKD is proposed to reinforce the clients' fitting of local data by introducing prior knowledge of local data distributions. Moreover, LKA is proposed to correct the distillation loss of the server, making the transferred local knowledge better match the generalized representation. Experiments on three datasets show that FedICT significantly outperforms all compared benchmarks in various data heterogeneous and model architecture settings, achieving improved accuracy with less than 1.2% training communication overhead compared with FedAvg and no more than 75% training communication round compared with FedGKT.

Motivation & Objective

• Address the challenge of training personalized, heterogeneous models in resource-constrained multi-access edge computing (MEC) environments.
• Overcome the limitations of existing federated multi-task learning (FMTL) methods that require high communication overhead and homogeneous model architectures.
• Eliminate reliance on public datasets in knowledge distillation for privacy-preserving training in MEC.
• Mitigate client drift caused by divergent local optimization directions in federated settings.
• Achieve high model accuracy with minimal communication cost and fast convergence in heterogeneous data and model settings.

Proposed method

• Introduce Federated Prior Knowledge Distillation (FPKD) to enhance clients' fitting of local data by leveraging prior knowledge of local data distributions.
• Propose Local Knowledge Adjustment (LKA) to correct the distillation loss at the server, aligning transferred local knowledge with the global model's generalized representation.
• Implement bidirectional knowledge distillation between clients and server, enabling mutual knowledge transfer without sharing raw data or model parameters.
• Use data distribution vectors at both client and server sides to guide knowledge transfer, ensuring alignment with local and global data characteristics.
• Design a two-phase training process: (1) client-side initialization with local data distribution vectors, and (2) server-side global model update using aggregated client knowledge.
• Apply both similarity-based and class-balanced LKA techniques to improve knowledge alignment under data skew.

Experimental results

Research questions

• RQ1 Can knowledge distillation be effectively applied in federated multi-task learning without relying on public datasets in MEC?
• RQ2 How can client drift be mitigated when clients optimize toward divergent local objectives in heterogeneous MEC environments?
• RQ3 To what extent can communication overhead be reduced in federated multi-task learning with heterogeneous model architectures?
• RQ4 Can personalized models be trained effectively under data and model heterogeneity while preserving privacy?
• RQ5 How does the proposed method compare in convergence speed and accuracy to state-of-the-art federated distillation frameworks?

Key findings

• FedICT achieves improved model accuracy across all tested datasets, including image classification and transportation mode detection, under diverse data and model heterogeneity.
• Communication overhead in FedICT is less than 1.2% of FedAvg’s overhead, significantly reducing bandwidth usage in MEC.
• FedICT reduces training rounds by up to 75% compared to FedGKT, accelerating convergence without sacrificing performance.
• The ablation study confirms that both FPKD and LKA contribute to performance gains, with degradation observed when either component is removed.
• FedICT outperforms FedGKT and FedDKC in average user accuracy while requiring fewer training rounds and less communication.
• Computation cost is manageable and outweighed by convergence speed gains, as additional overhead is orders of magnitude smaller than standard backpropagation.