Co-distillation-based defense framework for federated knowledge graph embedding against poisoning attacks

Yiqin Lu

doi:10.1016/j.csi.2025.104113

Co-distillation-based defense framework for federated knowledge graph embedding against poisoning attacks

2025-12-09

0

PRE

AI

DOI：10.1016/j.csi.2025.104113

Original

Original request for help

Save

Abstract

En 中文

• This study systematically defines and characterizes untargeted poisoning attack threats in federated knowledge graph embedding (FKGE), and pioneeringly proposes the CoDFKGE, the first co-distillation defense framework against poisoning attacks in FKGE. • By deploying customized bidirectional co-distillation mechanisms at the client side, CoDFKGE can not only effectively eliminate malicious operations in targeted attacks but also significantly resist performance degradation in untargeted attack scenarios. • Without requiring additional modifications to the baseline model, this method substantially enhances the defense capability against poisoning attacks while effectively reducing communication overhead, establishing a novel paradigm for secure federated knowledge graph learning.

Co-distillation-based defense framework for federated knowledge graph embedding against poisoning attacks

Abstract

Journal

Organization