Federated learning with differential privacy via fast Fourier transform for tighter-efficient combining

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Privacy

Background:

• Federated learning (FL) enables collaborative model training without centralizing sensitive data.
• Existing FL methods face privacy leakage risks, necessitating robust protection mechanisms.
• Balancing data privacy with model utility remains a key challenge in FL.

Purpose of the Study:

• To propose an improved Differential Privacy (DP) algorithm for enhancing privacy protection in federated learning.
• To optimize the computation of the privacy budget using Fast Fourier Transform (FFT) for efficiency.
• To improve the accuracy and stability of federated learning models under privacy constraints.

Main Methods:

• Implementation of an improved Differential Privacy (DP) algorithm tailored for federated learning.
• Utilization of Fast Fourier Transform (FFT) for efficient privacy budget computation.
• Adoption of Privacy Loss Distribution (PLD) and privacy curves for nuanced privacy analysis.
• Introduction of a novel temper sigmoid activation function to stabilize model training.

Main Results:

• The improved DP algorithm effectively tightens parameter bounds and reduces human influence on privacy.
• The use of FFT and PLD minimizes computational workload and errors from truncation/discreteness.
• The temper sigmoid activation function smooths accuracy curves and mitigates training fluctuations.
• Simulation results demonstrate a superior balance between privacy and utility in federated learning.

Conclusions:

• The proposed enhanced DP algorithm offers a more effective solution for privacy preservation in federated learning.
• The integration of FFT and advanced privacy analysis techniques improves computational efficiency and accuracy.
• The study successfully addresses the trade-off between privacy and utility, paving the way for more secure FL applications.