DSPFL: A Deep-Layer Sign Sharing Personalized Federated Learning Scheme for Mitigating Poisoning Attacks.

Area of Science:

• Cybersecurity
• Machine Learning
• Industrial Internet of Things (IIoT)

Background:

• Machine learning (ML) and federated learning (FL) are crucial for Industrial Internet of Things (IIoT) security.
• Poisoning attacks threaten FL models, especially with non-independent and identically distributed (non-IID) data in IIoT.
• Distinguishing malicious local models is challenging due to data heterogeneity.

Purpose of the Study:

• To propose a robust federated learning scheme resilient to poisoning attacks in IIoT environments.
• To enhance the accuracy and stability of personalized anomaly detection models in IIoT.
• To address the challenges posed by non-IID data and adversarial manipulations in FL.

Main Methods:

• Introduction of the deep-layer sign-sharing personalized FL (DSPFL) scheme.
• Aggregation of stochastic gradient signs (SignSGD) from deep layers of local models.
• Local retention of shallow model layers for personalization and privacy.

Main Results:

• DSPFL demonstrates enhanced robustness against poisoning attacks.
• The scheme improves the accuracy and stability of personalized local models.
• Experimental results show up to 20% higher overall personalized model accuracy compared to existing methods.

Conclusions:

• DSPFL effectively mitigates poisoning attacks in personalized FL for IIoT.
• The proposed method offers a promising solution for secure and accurate anomaly detection in smart industries.
• DSPFL balances global model robustness with local model personalization under adversarial conditions.