DefendFL: A Privacy-Preserving Federated Learning Scheme Against Poisoning Attacks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Federated learning (FL) enables collaborative model training without direct data sharing, but faces vulnerabilities to privacy leakage and poisoning attacks.
• Mask-based FL frameworks offer efficiency and functionality but are susceptible to sophisticated poisoning attacks.
• Existing methods lack practical detection mechanisms for poisoning within mask-based FL.

Purpose of the Study:

• To propose DefendFL, an efficient, privacy-preserving, and poisoning-detectable mask-based FL scheme.
• To address the vulnerability of mask-based FL to poisoning attacks.
• To provide practical means for detecting and mitigating malicious activities in FL.

Main Methods:

• Utilized a collinearity mask to safeguard gradient privacy.
• Employed cosine similarity for detecting malicious masked gradients.
• Implemented a mask verification mechanism to ensure aggregation validity and prevent mask manipulation.
• Resisted poisoning by removing or down-weighting malicious gradients during aggregation.

Main Results:

• DefendFL effectively detects and mitigates poisoning attacks in mask-based FL.
• The proposed scheme demonstrates superior efficiency compared to existing privacy-preserving detection methods.
• Security analysis and experimental evaluations validate the effectiveness of DefendFL.

Conclusions:

• DefendFL offers a robust solution for enhancing security in mask-based federated learning.
• The framework successfully balances privacy preservation with effective defense against poisoning attacks.
• DefendFL represents a significant advancement in securing collaborative machine learning environments.