MemberShield: A framework for federated learning with membership privacy.

Area of Science:

• Machine Learning
• Cybersecurity
• Data Privacy

Background:

• Federated Learning (FL) enables collaborative model training while preserving data privacy.
• FL systems are vulnerable to Membership Inference Attacks (MIA), threatening data confidentiality.
• Current MIA defenses often reduce model utility and increase computational overhead.

Purpose of the Study:

• To analyze the root causes of MIA vulnerability in FL, stemming from model overfitting.
• To propose MemberShield, a novel defense mechanism against MIA in FL.
• To evaluate MemberShield's effectiveness in privacy protection, model utility, and efficiency.

Main Methods:

• MemberShield employs a two-pronged approach: label preprocessing (one-hot to soft labels) and early stopping based on validation accuracy.
• The method focuses on reducing the differences in model behavior between member and non-member data samples.
• Empirical evaluations were conducted using three datasets and four distinct model architectures.

Main Results:

• MemberShield significantly outperforms existing state-of-the-art defenses against all forms of MIA.
• The proposed method achieves superior privacy protection while maintaining high model utility.
• MemberShield demonstrates a substantial reduction in training time and is easy to implement.

Conclusions:

• MemberShield offers a practical and effective solution for enhancing privacy in Federated Learning.
• The generalization-based defense successfully addresses MIA vulnerabilities without compromising model performance.
• This approach provides a scalable and efficient method for secure collaborative machine learning.