PLDP-FL: Federated Learning with Personalized Local Differential Privacy.

Area of Science:

• Machine Learning
• Data Privacy
• Cybersecurity

Background:

• Federated learning (FL) addresses data silos and privacy but offers insufficient protection.
• Existing secure FL schemes using local differential privacy (LDP) lack personalized privacy settings.
• Uniform privacy budgets in LDP can lead to inadequate or excessive privacy for different clients.

Purpose of the Study:

• To introduce a perturbation algorithm (PDPM) for personalized local differential privacy (PLDP).
• To enable clients to adjust privacy parameters based on data sensitivity.
• To provide tailored privacy protection in federated learning environments.

Main Methods:

• Development of a perturbation algorithm (PDPM) for personalized local differential privacy (PLDP).
• Implementation of client-adjustable privacy parameters based on data sensitivity.
• Rigorous privacy proof and simulation on synthetic and real datasets.

Main Results:

• The proposed PDPM algorithm successfully achieves personalized privacy protection.
• Experiments demonstrate high-quality model generation with PLDP.
• The scheme effectively addresses the issue of uniform privacy budgets in federated learning.

Conclusions:

• The PDPM algorithm offers a novel solution for personalized privacy in federated learning.
• This approach enhances data privacy without compromising model utility.
• The findings support the adoption of PLDP for more equitable and effective federated learning.