FedLAD: A Linear Algebra Based Data Poisoning Defence for Federated Learning

EasyChair Preprint 16009

Abstract

Sybil attacks threaten federated learning since malicious nodes can collude and occupy the majority, hence overwhelm the federated learning system. It is imperative to develop a countermeasure to ensure the security of federated learning systems. We introduce a novel data poisoning (one of the Sybil attacks) defence method for federated learning, named Linear Algebra based Detection (FedLAD). Unlike existing approaches such as clustering and robust-training which struggle in scenarios where malicious nodes occupy the majority, FedLAD models federated learning aggregation as a linear problem and turn it into a linear algebra optimisation problem. This approach identifies potential attacks by finding the independent linear combination from its original linear combination which is the simplest form of the original linear combination. It contains only the important elements with redundant and malicious elements filtered out. Extensive experimental evaluations demonstrate the efficacy of FedLAD compared to five well-established defence methods: Sherpa, CONTRA, Median, Trimmed Mean, and Krum. Using tasks from both image classification and natural language processing, the experiments confirm that FedLAD is robust and independent of specific application settings. Results show that FedLAD effectively safeguards federated learning systems across a wide range of malicious node ratios. Compared with baseline defence methods, FedLAD maintains a low attack success rate for malicious nodes when their ratio ranges from 0.2 to 0.8. Additionally, it preserves high model accuracy when the malicious node ratio is between 0.2 and 0.5. These findings highlight FedLAD’s potential to enhance the reliability and performance of Federated Learning systems with data poisoning attacks.

Keyphrases: Data Poisoning Detection, Federated Learning, machine learning

@booklet{EasyChair:16009,
  author    = {Qi Xiong and Hai Dong and Nasrin Sohrabi and Zahir Tari},
  title     = {FedLAD: A Linear Algebra Based Data Poisoning Defence for Federated Learning},
  howpublished = {EasyChair Preprint 16009},
  year      = {EasyChair, 2026}}