ALDP-FL for adaptive local differential privacy in federated learning

Area of Science:

• Machine Learning
• Data Privacy
• Cybersecurity

Background:

• Federated learning (FL) trains models on decentralized data without sharing raw user information.
• However, FL is vulnerable to privacy attacks where attackers infer sensitive data from model updates.
• Existing privacy-preserving methods in FL often struggle to balance privacy guarantees with model utility.

Purpose of the Study:

• To propose a novel privacy-preserving federated learning method, Adaptive Localized Differential Privacy Federated Learning (ALDP-FL).
• To enhance data privacy in federated learning by dynamically injecting noise into model updates.
• To maintain high model accuracy while providing robust protection against sensitive information inference.

Main Methods:

• ALDP-FL dynamically adjusts the clipping threshold for each network layer based on the moving average of their norm.
• Adaptive noise is injected into each layer, tailored to the specific characteristics of the updates.
• A bounded perturbation mechanism is employed to mitigate the accuracy degradation caused by the added noise.

Main Results:

• ALDP-FL demonstrated significant improvements across key metrics: Accuracy (+10.57%), Precision (+10.64%), Recall (+10.52%), and F1 Score (+10.64%).
• The method showed superior performance in defending against iDIG attack reconstruction, with MSE improving by 391.2% and SSIM by -85.4%.
• Experiments on MNIST, Fashion MNIST, and CIFAR-10 datasets validated the effectiveness and practicality of ALDP-FL.

Conclusions:

• ALDP-FL effectively enhances privacy in federated learning without substantial loss of model performance.
• The adaptive noise injection and bounded perturbation mechanisms offer a robust solution for privacy-preserving machine learning.
• The proposed method significantly outperforms existing techniques in protecting user data against sophisticated inference attacks.