FedAdOb: Privacy-Preserving Federated Deep Learning with Adaptive Obfuscation

Area of Science:

• Machine Learning
• Data Privacy
• Cybersecurity

Background:

• Federated learning (FL) enables collaborative model training without data sharing, but privacy leakage remains a concern.
• Existing defense mechanisms often compromise model performance due to fixed data or gradient obfuscation.
• Attacks on FL privacy have spurred research into robust defending strategies.

Purpose of the Study:

• To propose FedAdOb, a novel adaptive obfuscation mechanism for enhanced data privacy in federated learning.
• To protect private features and labels without degrading original model performance.
• To ensure privacy preservation in both horizontal and vertical federated learning scenarios.

Main Methods:

• Introduction of FedAdOb, a passport-based adaptive obfuscation technique.
• Theoretical validation of FedAdOb's privacy-preserving capabilities for features and labels (Theorems 1 and 2).
• Extensive experimental evaluations on diverse datasets and network architectures.

Main Results:

• FedAdOb effectively protects private data in federated learning settings.
• The proposed mechanism achieves a superior trade-off between privacy preservation and model performance.
• Experimental results demonstrate FedAdOb's superiority over existing privacy-preserving methods.

Conclusions:

• FedAdOb offers a robust solution for privacy-preserving federated learning.
• The adaptive obfuscation strategy maintains high model performance while ensuring data confidentiality.
• FedAdOb represents a significant advancement in securing collaborative machine learning.