Secure and Efficient Federated Learning Against Model Poisoning Attacks in Horizontal and Vertical Data Partitioning

Area of Science:

• Distributed Systems
• Machine Learning Security

Background:

• Hybrid federated learning (FL) combines horizontal and vertical data partitioning for enhanced distributed systems.
• Hybrid FL is susceptible to model poisoning attacks, compromising global model integrity.
• Existing defenses face high detection costs and accuracy issues due to data diversity.

Purpose of the Study:

• To develop a secure and efficient hybrid FL framework against model poisoning attacks.
• To minimize training costs and energy consumption while maintaining accuracy.
• To address the challenges of detecting malicious updates in diverse FL environments.

Main Methods:

• Defined two novel local model poisoning attacks for analysis.
• Analyzed execution time and energy consumption in hybrid FL.
• Formulated an optimization problem solved via Markov decision process and multiagent reinforcement learning (MARL).
• Proposed a malicious device detection (MDD) method using MARL.
• Introduced a poisoned model detection (PMD) method based on model change consistency.

Main Results:

• The MDD method reduced training costs by over 50% against random local model poisoning attacks.
• Combined MDD and PMD methods maintained desired accuracy under advanced adaptive local model poisoning (ALMP) attacks.
• Both methods demonstrated reductions in execution time and energy consumption.

Conclusions:

• The proposed MARL-based MDD and PMD methods offer a robust defense against hybrid FL model poisoning.
• The approach effectively balances security, efficiency, and accuracy in decentralized learning environments.
• This work provides a significant advancement in securing hybrid federated learning systems.