A privacy preservation framework for feedforward-designed convolutional neural networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Feedforward-designed convolutional neural networks (FF-CNNs) offer interpretable models with low training complexity.
• Current FF-CNNs achieve high performance in image classification and point cloud processing.
• Existing privacy-preserving methods are incompatible with FF-CNN training, posing a user privacy risk.

Purpose of the Study:

• To analyze and address the inherent risk of user privacy leakage in FF-CNN training.
• To propose a novel algorithm, securely forward-designed convolutional neural network (SFF-CNN), for enhanced data privacy.
• To ensure data provider security within the FF-CNN framework.

Main Methods:

• Introduction of the DPSaab algorithm for noise addition during the Saab transform in FF-CNN design.
• Implementation of the SJS algorithm to filter features for the fully connected layer, mitigating overfitting and privacy risks.
• Theoretical proof of differential privacy and experimental validation of the SFF-CNN algorithm.

Main Results:

• The SFF-CNN algorithm effectively protects user privacy during FF-CNN model training.
• The proposed DPSaab and SJS algorithms demonstrate strong privacy protection capabilities.
• SFF-CNN outperforms existing deep learning privacy-preserving algorithms in utility and robustness.

Conclusions:

• The developed SFF-CNN algorithm provides a secure and effective solution for privacy preservation in FF-CNN models.
• The study highlights the vulnerability of FF-CNNs to privacy leakage and offers a robust countermeasure.
• SFF-CNN represents a significant advancement in secure deep learning for sensitive data applications.