Deep Neural Network Quantization Framework for Effective Defense against Membership Inference Attacks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Machine learning on edge devices faces computational and privacy challenges.
• Membership inference attacks (MIA) threaten user data privacy by inferring training set membership.
• Protecting training data is crucial for privacy-sensitive applications like healthcare.

Purpose of the Study:

• To investigate the impact of quantization on privacy leakage.
• To propose a novel quantization method specifically designed to enhance resistance against MIA.
• To develop a defense mechanism against MIA for neural networks.

Main Methods:

• Exploiting the implications of quantization on privacy leakage.
• Developing and proposing a novel quantization framework prioritizing MIA resistance.
• Evaluating the proposed method on benchmark datasets and various model architectures.

Main Results:

• The proposed quantization method enhances neural network resistance to MIA.
• Experimental results show improvements in precision, recall, and F1-score compared to full bitwidth models.
• For ResNet on Cifar10, MIA attack accuracy was reduced by 14%, true positive rate by 37%, and F1-score of members by 39%.

Conclusions:

• Quantization can be leveraged to bolster privacy against MIA.
• The proposed quantization framework effectively defends against MIA without compromising performance.
• This method offers a viable solution for enhancing data privacy in machine learning deployments.