Novel Constructions for Computation and Communication Trade-offs in Private Coded Distributed Computing
Shanuja Sasi, Onur Günlü
TL;DR
The paper tackles privacy in distributed computation by enforcing task-index privacy in a MapReduce-like CDC setting. It introduces an extended PDA framework that combines two PDAs to produce private CDC schemes, and derives two families of extended PDAs to analyze computation-communication trade-offs under privacy. The first construction preserves the same computation load as non-private CDC but scales the communication load by a factor $Q$, while the second construction yields improved trade-offs by trading some computation for reduced communication, with explicit load expressions showing the privacy penalties and gains. The work blends ideas from coded caching, PDAs, and distributed computing to provide information-theoretic privacy for task assignments without encrypting data, offering a path toward practical, privacy-aware private CDC designs albeit with scalability challenges. The results quantify how privacy constraints impact communication, and propose PDA-based strategies to mitigate overhead and achieve near-optimal private CDC performance in decentralized settings.
Abstract
Distributed computing enables scalable machine learning by distributing tasks across multiple nodes, but ensuring privacy in such systems remains a challenge. This paper introduces a novel private coded distributed computing model that integrates privacy constraints to keep task assignments hidden. By leveraging placement delivery arrays (PDAs), we design an extended PDA framework to characterize achievable computation and communication loads under privacy constraints. By constructing two classes of extended PDAs, we explore the trade-offs between computation and communication, showing that although privacy increases communication overhead, it can be significantly alleviated through optimized PDA-based coded strategies.