Security and Efficiency Enhancement for Split Learning: A Machine Learning based Malicious Clients Detection Approach

Ensuring data privacy and mitigating the potential impact of malicious clients are crucial considerations in split learning frameworks. To improve the security and efficiency of split learning, we propose a novel approach that integrates the hashcash algorithm with a deep neural network based malicious client detection mechanism. Specifically, we first design a supervised deep neural network based algorithm to detect potentially malicious clients by analyzing an array of client attributes, such as IP address, region, and behavioral patterns. This serves as a risk ranking system to evaluate each client, indicating the likelihood of malicious behavior. Subsequently, we design a hashcash based algorithm to verify the clients’ legitimacy and computing capabilities for completing the split learning task, using puzzles with different difficulty levels. By using open source datasets and machine learning models, the simulation results demonstrate the effectiveness of the proposed method in identifying and mitigating the impact of malicious clients while enhancing the overall security and efficiency of split learning scenarios.