Privacy-preserving Online Federated Learning for Massive Infinite Streams

Area of Science:

• Computer Science
• Machine Learning
• Data Privacy

Background:

• Online federated learning (OFL) is crucial for privacy-preserving analytics on decentralized data streams.
• OFL faces unique challenges like longitudinal privacy leakage and accumulated costs due to infinite data streams, unlike batch-based FL.

Purpose of the Study:

• To extend differential privacy (DP) to OFL for window-based privacy protection on infinite streams.
• To propose a generic sampling-based framework for DP-enhanced OFL algorithms.
• To develop an adaptive sampling strategy for improved utility and efficiency in OFL.

Main Methods:

• Extended traditional differential privacy (DP) to OFL with tunable granularity for infinite streams.
• Developed a generic sampling-based solution framework for DP-enhanced OFL.
• Introduced Sampling³-OFL, an adaptive triple-sampling strategy using deep reinforcement learning.

Main Results:

• The sampling solution framework achieves asymptotic optimality under DP constraints for infinite time horizons.
• Sampling³-OFL dynamically determines near-optimal sampling strategies.
• Experiments show Sampling³-OFL scales to millions of streams, improving utility by 0.74%-15.84% and reducing communication costs by 33.33%-95.24%.

Conclusions:

• The proposed DP-enhanced OFL framework with adaptive sampling offers significant improvements in utility and efficiency.
• Sampling³-OFL provides a scalable and effective solution for privacy-preserving online collaborative analytics.
• The adaptive triple-sampling strategy demonstrates the potential of deep reinforcement learning in optimizing OFL performance.