Spiking Neural Networks and Adversarial Attack Project

---

Overview

This repository contains Python code for the paper titled "Neuromorphic Computing Paradigms Enhance Robustness through Spiking Neural Networks."

The project investigates how SNN can improve robustness using temporal processing capabilities.

This repository includes codes for:

• Development of an SNN architecture.
• Evaluation of SNN under adversarial attack.
• Analysis of SNN robustness.

Installation

Prerequisites

Ensure you have Python 3.8 or later installed. Install the required packages using the following command:

pip install -r requirements.txt

Recommend Libraries

• torch >= 2.0.0+cu118
• torchvision >= 0.15.1+cu118
• numpy >= 1.24.2
• matplotlib >= 3.7.1
• tqdm >= 4.61.2
• umap-learn == 0.5.7
• gdown == 5.2.0

You can also run with the following Libraries

• pytorch == 2.4.0 # on cpu
• torchvision == 0.19.0
• numpy >= 1.24.2
• matplotlib >= 3.7.1
• tqdm >= 4.61.2
• gdown == 5.2.0

To install PyTorch, refer to the official PyTorch installation guide.

Typical install time on a "normal" desktop computer depends on the network condition!

Project Structure

part1 corresponds to the results in Prioritizing task-critical temporal information during encoding boosts SNN robustness

part2 corresponds to the results in Training algorithms for SNNs enabling flexible encoding achieve robust generalization

part3 corresponds to the results in Early exiting SNN decoding protects against subsequent perturbations improving robustness

part4 corresponds to the results in SNN employs fused input encoding to enhance performance in complex attacks

|-- global_tools/           # Tools for adversarial attack and for training SNNs
|-- global_configs.py            # Script for setting dataset, and encoding methods, as well as url for model checkpoints
|-- data/                   # place to save clean datasets

|-- part1/                  # Codes evaluating the robustness of SNN with regard to encoding methods
|-- part2/                  # Codes evaluating the robustness of SNN with regard to training methods
|-- part3/                  # Codes evaluating the robustness of SNN with regard to decoding methods
|-- part4                   # Codes evaluating the scability of the proposed methods
|-- part4_attack/           # place to save attacked CIFAR-10 dataset

|-- requirements.txt        # Python dependencies
|-- README.md               # Project documentation

Usage

Before you run

Go to global_tools/global_configs and set

dataset_path = {
    'mnist': your_data_path/mnist,
    'fashionmnist': your_data_path/fashionmnist,
    'cifar10': your_data_path/cifar10
}

1. Download checkpoints

To download pretrained checkpoints for evaluation, run

python download_all.py

2. Evaluate SNNs/ANNs in part1, part2, part3, part4

To get all data to plot the main figure corresponding to the main results, go to the corresponding directories and run:

python run_all_eval.py

Expected run time on a "normal" desktop computer: within 10 seconds.

2. Plots in supplementary materials

To get all data to plot the figures in the supplementary materials, go to the corresponding directories and run:

python run_supp_fig.py

In case one want to train the model:

1. Obtaining the SNN/ANN Model in part1

To train the ANN model in part1, go to part1_tools and run

python run_ANN_train.py -dataset=MNIST -hidden=100 -bb=0 -seed=A

Then, to get SNN, run

python run_SNN_convert.py -dataset=MNIST -hidden=100 -pth_file=model_MNIST_seedA.pth

2. Obtaining the SNN/ANN Model in part2, part3, part4

To train the SNN/ANN model in partx, go to partx_tools and run

python run_train.py -config=../net_params/config_name

config_name is a placeholder for YAML files in net_params.

Then, to get converted SNN for part2, open run_SNN_convert.py in part2_tools, change seed, and then run

python run_SNN_convert.py

3. Obtaining the SNN Model with fused encoding in part4

Find configs that have string 'mix' in the name, such as CIFAR10_ResNet18_ACTBP_mix_entropy-CRT.yaml.

Then, to get SNN, go to part4_tools and run

python run_train-mix.py -config=../net_params/config_name

License

This project is licensed under the MIT License. See LICENSE for details.