Related Experiment Video
Updated: May 10, 2025
11:18
Closed-loop Neuro-robotic Experiments to Test Computational Properties of Neuronal Networks
Published on: March 2, 2015
10.2K
Encrypted Spiking Neural Networks Based on Adaptive Differential Privacy Mechanism
Xiwen Luo1,2, Qiang Fu1,2, Junxiu Liu1,2
1Guangxi Key Lab of Brain-Inspired Computing and Intelligent Chips, School of Electronic and Information Engineering, Guangxi Normal University, Guilin 541004, China.
Entropy (Basel, Switzerland)
|April 26, 2025
Summary
Adaptive Differential Private Spike Neural Networks (ADPSNN) dynamically adjust privacy budgets for improved efficiency and personal privacy. This approach enhances accuracy in image classification tasks using various spiking neuron models.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Computational Neuroscience
Background:
- Spike Neural Networks (SNNs) offer high performance but face challenges with differential privacy (DP) protocols.
- DP in SNNs often introduces uniform noise, potentially compromising the balance between model efficiency and user privacy.
Purpose of the Study:
- To introduce the Adaptive Differential Private SNN (ADPSNN) for dynamic privacy budget adjustment.
- To improve the trade-off between model utility and personal privacy in SNNs.
Main Methods:
- ADPSNN dynamically adjusts the privacy budget based on output spike and label correlations.
- Noise is applied to gradient parameters according to the adjusted privacy budget.
- Tested on MNIST, Fashion-MNIST, CIFAR10, and CIFAR100 datasets using Leaky Integrate-and-Fire (LIF) and Integrate-and-Fire (IF) neuron models.
Main Results:
- LIF neurons achieved 99.56% accuracy on MNIST and 92.26% on Fashion-MNIST.
- IF neurons achieved 90.67% accuracy on CIFAR10 and 66.10% on CIFAR100.
- ADPSNN improved accuracy by 0.09% to 3.1% compared to existing methods.
Conclusions:
- ADPSNN offers a superior approach to privacy-preserving SNNs.
- The choice of neuron model (LIF or IF) impacts performance across different datasets.
- ADPSNN shows promise for applications in image classification, healthcare, and intelligent driving.
