Utility-Leakage Trade-Off for Federated Representation Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Federated representation learning (FRL) enables decentralized data analysis without raw data sharing.
• Existing methods like differential privacy (DP) offer general protection but can degrade performance.
• A critical need exists to protect specific sensitive information while preserving utility.

Purpose of the Study:

• To investigate information-theoretic protection for sensitive attributes in FRL.
• To develop a method that balances utility and sensitive information leakage.
• To provide a tunable privacy-utility trade-off mechanism.

Main Methods:

• Utilizing mutual information to quantify utility and sensitive information leakage.
• Proposing a novel FRL method incorporating local DP.
• Maximizing utility under a constraint on sensitive information leakage (less than ϵ).

Main Results:

• The proposed scheme achieves superior utility-leakage trade-offs compared to baseline methods.
• The method effectively protects specific sensitive information (e.g., race).
• Controlling noise levels in local DP allows adjustable privacy-utility trade-offs.

Conclusions:

• The developed method offers a targeted approach to privacy in FRL.
• This technique enhances data utility while mitigating specific privacy risks.
• The tunable nature of the method provides flexibility for various applications.