Uldp-FL: Federated Learning with Across-Silo User-Level Differential Privacy

Area of Science:

• Computer Science
• Machine Learning
• Cryptography

Background:

• Differentially Private Federated Learning (DP-FL) is crucial for collaborative machine learning with formal privacy guarantees.
• Existing DP-FL methods typically ensure record-level privacy within silos for cross-silo FL.
• The challenge of achieving user-level DP when a single user's data spans multiple silos is an open research question.

Purpose of the Study:

• To propose Uldp-FL, a novel federated learning (FL) framework providing user-level differential privacy (DP) in cross-silo settings.
• To address scenarios where individual user data is distributed across multiple data silos.
• To establish a new standard for privacy in collaborative machine learning with distributed user data.

Main Methods:

• Developed Uldp-FL, a framework ensuring user-level DP through per-user weighted clipping, distinct from group-privacy methods.
• Conducted theoretical analysis of the algorithm's privacy and utility trade-offs.
• Implemented an enhanced weighting strategy based on user record distribution to improve utility.
• Designed a novel private protocol to prevent information leakage to silos and the server.

Main Results:

• Demonstrated substantial improvements in privacy-utility trade-offs under user-level DP compared to baseline methods.
• Experimental results on real-world datasets validate the effectiveness of the proposed Uldp-FL framework.
• The novel private protocol successfully prevented additional information disclosure.

Conclusions:

• Uldp-FL is the first FL framework to effectively provide user-level DP in the general cross-silo FL setting.
• The proposed method offers a significant advancement in balancing privacy and utility for distributed machine learning.
• This work sets a new benchmark for privacy guarantees in federated learning systems with distributed user data.