Decentralized Pliable Index Coding For Federated Learning In Intelligent Transportation Systems

TL;DR

This paper introduces decentralized pliable index coding solutions to improve data shuffling in federated learning for intelligent transportation systems, addressing non-IID data issues and enhancing model convergence and accuracy.

Contribution

It proposes novel decentralized pliable index coding schemes, including optimal solutions for specific cases, to facilitate efficient data shuffling in federated learning without a central server.

Findings

Improved convergence speed and accuracy of FL models.

Enhanced data distribution towards IID through proposed coding schemes.

Validated improvements on FedAvg and CELL techniques.

Abstract

Federated Learning is a promising option for data privacy and security in ITS, because it allows edge devices, Road Side Units (RSUs), and Central Server (CS) to jointly train the machine learning model. Since RSU collects data from the vehicles passing through its range, the local data of each RSU will have a non-IID distribution, which adversely affects the convergence speed and accuracy of FL training. Generating synthetic data locally at individual nodes, followed by data shuffling among the nodes, is a promising approach to address the Non-IID data problem. In this work, we propose pliable index coding (PIC) solutions for efficient data shuffling among the nodes in an FL system. In PIC($S$) problems, a client is satisfied if it can retrieve any $S$ new messages not originally present in its side-information. We particularly consider decentralized pliable index coding problems (DPIC) where the clients communicate among themselves without a central server to model the data shuffling in FL. A class of DPIC, known as Consecutive Decentralized Pliable Index Coding (CDPIC($S$,$K$)), where each client has $K$ consecutive messages as side-information, is considered. For CDPIC($S$,$K$) problems, pliable index code designs are provided for any value of $K$ and $S$, and optimality proofs for some of the cases are established. Further, these CDPIC solutions are applied for data shuffling in FL, to transform the local data distribution towards IID progressively with each transmission, thereby enhancing the performance of FL. The improvement in the accuracy and convergence of the most popular FL technique, FedAvg, and a promising federated submodel technique, CELL (Communication Efficient Lottery Learning), are analysed by providing different degrees of data shuffling using the proposed CDPIC schemes.