Blockchain poisoning attacks in federated learning

Author

Mohammad Raihan Uddin

Date of Award

2025

Document Type

Thesis

Degree Name

Master of Science in Engineering (MSE)

Department

Electrical and Computer Engineering

Committee Chair

Dinh C. Nguyen

Committee Member

Earl Wells

Committee Member

David Coe

Research Advisor

Dinh C. Nguyen

Subject(s)

Federated learning (Machine learning), Artificial intelligence--Security measures, Privacy-preserving techniques (Computer science)

Abstract

Federated Learning (FL) enables collaborative model training across distributed clients without sharing raw data, thereby preserving privacy. However, conventional FL depends on a centralized aggregator, exposing it to poisoning and backdoor attacks that threaten model integrity. This thesis introduces B-ZkFed, a blockchain-based and zero-knowledge proof (ZKP)-enhanced FL framework designed to ensure verifiable trust, transparency, and robustness. The blockchain layer decentralizes aggregation through smart contracts, while the ZKP module allows clients to cryptographically prove the correctness of updates without revealing private data. A multi-layer adaptive defense combining gradient-norm filtering, model-similarity analysis, and robust aggregation mitigates poisoning threats under both IID and Non-IID settings. Experiments on the CIFAR-10 dataset show that B-ZkFed reduces attack success rates by up to 30.7% and sustains accuracy within 3-4% of the clean baseline, with minimal computational overhead. Overall, B-ZkFed establishes a verifiable, privacy-preserving, and tamper-resistant foundation for secure federated intelligence.

Recommended Citation

Uddin, Mohammad Raihan, "Blockchain poisoning attacks in federated learning" (2025). Theses. 793.
https://louis.uah.edu/uah-theses/793