Dynamic Asynchronous Anti Poisoning Federated Deep Learning with Blockchain-Based Reputation-Aware Solutions

Area of Science:

• Machine Learning
• Cybersecurity
• Distributed Systems

Background:

• Federated learning (FL) is a privacy-preserving machine learning technique enabling collaborative model training across decentralized clients.
• Existing FL frameworks face challenges with training inefficiency and vulnerability to data or model poisoning attacks.
• These limitations hinder the practical application and performance of federated learning systems.

Purpose of the Study:

• To propose a novel federated deep learning framework that addresses both efficiency and security concerns in FL.
• To enhance the speed of model averaging and improve the overall training efficiency.
• To develop robust defenses against various poisoning attacks targeting federated learning models.

Main Methods:

• Introduced a lightweight dynamic asynchronous algorithm with adaptive straggler removal based on computing power, channel conditions, and parameter anomalies.
• Implemented a novel local reliability mutual evaluation mechanism for detecting anomalous parameters indicative of poisoning attacks.
• Adjusted the weight proportion during model aggregation based on the calculated evaluation scores to mitigate attack impact.

Main Results:

• The proposed framework significantly reduces training time by up to 30% compared to existing methods.
• Demonstrated robust performance against representative poisoning attacks, maintaining model integrity and accuracy.
• Experimental results on three datasets validated the effectiveness and applicability of the designed framework.

Conclusions:

• The dynamic asynchronous anti-poisoning federated deep learning framework effectively improves FL efficiency and security.
• The proposed methods successfully address straggler issues and enhance resilience against poisoning attacks.
• This approach offers a practical solution for deploying secure and efficient federated learning systems.