A defense method against multi-label poisoning attacks in federated learning.

Area of Science:

• Machine Learning
• Cybersecurity
• Data Privacy

Background:

• Federated learning (FL) enables collaborative model training without raw data sharing, enhancing privacy.
• FL systems are susceptible to data poisoning attacks, such as label flipping, compromising model integrity.
• Existing defenses struggle against sophisticated multi-label flipping attacks.

Purpose of the Study:

• To propose a robust defense mechanism against multi-label flipping attacks in federated learning.
• To enhance the security and reliability of federated learning models against sophisticated adversarial manipulations.

Main Methods:

• Extracting gradients from output layer neurons.
• Applying clustering analysis using a combination of metrics to differentiate benign and malicious participants.
• Identifying and filtering out malicious model updates.

Main Results:

• The proposed method effectively identifies and filters malicious updates.
• Demonstrates strong robustness against multi-label flipping attacks, even with a high proportion of attackers.
• Outperforms existing defense methods in accuracy and robustness across diverse datasets (MNIST, FashionMNIST, NSL-KDD, CICIDS-2017).

Conclusions:

• The developed defense method offers significant improvements in securing federated learning against advanced poisoning attacks.
• The approach provides a reliable solution for maintaining model integrity in privacy-preserving distributed learning environments.
• Effective against varied attack scenarios and a high percentage of malicious participants.