Colony: Robust Personalized ML With Contribution-Driven Connectivity and Resource Allocation in P2P Networks

Abstract

Decentralized Personalized Learning (DPL) is an emerging paradigm that enables nodes to train personalized models across a peer-to-peer network without sharing their data or relying on a central server. However, nodes are often connected to a dynamic set of neighbors without prior knowledge of the quality of their datasets, reliability, communication efficiency, or adversarial behavior. To ensure convergence with minimal resource usage, prior work has often relied on centralized optimization via a coordinator (e.g., base station), which requires substantial computation, especially in dynamic networks at scale. In this work, we formulate DPL as a cooperative game in which nodes seek to maximize learning utility subject to resource budgets. To this end, we introduce Colony, a fully decentralized framework that (i) assigns contribution scores based on neighbors' estimated Shapley values, driven by a utility function that penalizes harmful updates, and (ii) computes an efficiency score that combines link reliability, latency, and energy for resource-aware neighbor selection. These scores jointly enable nodes to optimize the effectiveness and efficiency of their neighborhoods each training round, eliminating the need for heavy per-round global optimization. Our experiments on MNIST and CIFAR-10 demonstrate that Colony preserves personalization and robustness in the presence of noisy or Byzantine neighbors while significantly reducing bandwidth and energy consumption compared with topology-agnostic baselines. Concretely, Colony maintains a stable convergence rate even with 45% Byzantine nodes, achieving up to a threefold reduction in transmission energy, which can be further adjusted by each node according to its energy and bandwidth budgets.