Addressing unreliable local models in federated learning through unlearning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Federated learning (FL) systems face challenges in maintaining global model reliability due to unreliable local models.
• Existing federated unlearning (FUL) methods primarily address bad data, neglecting other negative influence sources like adversarial attacks or communication constraints.

Purpose of the Study:

• To introduce Local Model Refining (LMR), a novel FUL method designed to mitigate negative impacts from both bad data and other factors.
• To enhance the reliability and accuracy of global models in federated learning systems.

Main Methods:

• LMR categorizes unreliable local models based on influence source: bad data or other factors.
• Bad Data Influence Unlearning (BDIU) is a client-side algorithm using gradient ascent to mitigate bad data effects.
• Other Influence Unlearning (OIU) is a server-side algorithm that reconstructs local models using previous global model parameters.

Main Results:

• LMR effectively identifies and mitigates negative influences from diverse sources.
• Evaluations on MNIST, FMNIST, CIFAR-10, and CelebA datasets demonstrate enhanced accuracy.
• LMR achieves an average unlearning speedup of 5x compared to existing methods.

Conclusions:

• LMR offers a comprehensive solution for federated unlearning by addressing multiple sources of unreliability.
• The proposed method significantly improves global model performance and unlearning efficiency in federated learning.