DOCAttack: Data-Oriented Community Attack in Decentralized Federated Learning

doi:10.1109/tnse.2026.3702609

DOCAttack: Data-Oriented Community Attack in Decentralized Federated Learning

2026-06-11

0

PRE

AI

DOI：10.1109/tnse.2026.3702609

Original

Original request for help

Save

Abstract

En 中文

Decentralized federated learning (DFL) relies on peer-to-peer communication, making its convergence highly sensitive to how non-independent and identically distributed (non-IID) client data propagate across the network. Existing structural attacks operate purely on topology and therefore miss the vulnerabilities that stem from the interaction between non-IID client data and the network structure. In this paper, we propose a data-oriented community attack (DOCAttack) framework in DFL. First, the proposed framework fuses leaked client data profiles with network structure to learn data-aware node embeddings. Then, it detects data-oriented communities from these embeddings. Finally, it optimizes node and edge removals under a given budget. Experimental results show that DOCAttack produces stronger structural disruption than topology-only baselines and consistently isolates data-oriented communities that these baselines fail to separate. In addition, these disruptions prevent the system from integrating information across communities, leading to degraded learning performance and the collapse of global consensus.

Keywords:

Decentralized federated learning

adversarial attack

network dismantling

data-oriented communities

graph representation learning

DOCAttack: Data-Oriented Community Attack in Decentralized Federated Learning

Abstract

Journal

Organization