Threshold Switching Memristor-Based Radial-Based Spiking Neuron Circuit for Conversion Based Spiking Neural Networks Adversarial Attack Improvement

The analog neural network to spiking neural network (ANN-to-SNN) conversion is an effective method for improving the performance of SNNs. However, the existing mainstream conversion method (rectified linear unit, ReLU) still face the problem of weak ability for adversarial attacks. In this work, ins...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on circuits and systems. II, Express briefs Express briefs, 2024-03, Vol.71 (3), p.1-1
Main Authors: Wu, Zuheng, Li, Wei, Zou, Jianxun, Feng, Zhe, Chen, Tao, Fang, Xiuquan, Li, Xing, Zhu, Yunlai, Xu, Zuyu, Dai, Yuehua
Format: Article
Language:English
Subjects:
Adversarial attacks
Analog circuits
ANN-to SNN
Biological properties
Biological system modeling
Biology
Computational modeling
Integrated circuit modeling
Memristor
Memristors
Neural networks
Neurons
Radial basis function
Radial basis spiking neuron (RBSN) circuit
Silicon nitride
Spiking
Switching
Threshold voltage
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The analog neural network to spiking neural network (ANN-to-SNN) conversion is an effective method for improving the performance of SNNs. However, the existing mainstream conversion method (rectified linear unit, ReLU) still face the problem of weak ability for adversarial attacks. In this work, inspired by the radial basis function and the "near enhancement and far inhibition (NEFI)" properties of biological neurons, a threshold switching (TS) memristor based radial basis spiking neuron (RBSN) circuit is proposed for the ANN-to-SNN conversion implementation. The results indicate that the RBSN circuit can effectively implement the NEFI spiking properties, which is benefit for filtering the adversarial attacks information. Furthermore, the comparison of ReLU and RBSN conversion methods based ANN-to-SNN for MNIST dataset classification task was performed. The results indicate that the RBSN shows obviously advantage than the ReLU for confronting the adversarial attacks. The accuracy of RBSN based ANN-to-SNN achieved ∼80.6%, even the input data containing 40% attack information, whereas the ReLU based ANN-to SNN only achieved ∼49.2%. This work provides new ideas for the security design of neuromorphic computing systems.
ISSN:1549-7747
1558-3791
DOI:10.1109/TCSII.2023.3318592