snn_attack

This code apllies gradient based SNN attack on MNIST dataset

Train an SNN model

CUDA_VISIBLE_DEVICES=0 python train_mnist_ce.py

Collect samples from each classes (optional)

We have already selected 5 images for each class and saved these samples in file ckpt. The following command is used to select data.

python dataset_select.py

Apply SNN attack

CUDA_VISIBLE_DEVICES=0 python snn_attack.py

In file 'snn_attack.py', 'attack_mode' indicates targeted or untargeted attack, 'thresh_attack_value' controls the firing threshold in pernultimate layer.

For the boundary attack, please see the original source code! for more detail. Note that the boundary attack does not have any constrain on the model.

Results

untargeted attack; pernultimate layer firing thresh = original firing threshold = 0.3
'succ': 46, 'fail': 4, 'wrong': 0, 'ratio': 0.92, 'avg_diff': 0.024219149865372026

untargeted attack; pernultimate layer firing thresh = 1.0
'succ': 50, 'fail': 0, 'wrong': 0, 'ratio': 1.00, 'avg_diff': 0.020177705427631736

targeted attack; pernultimate layer firing thresh = original firing threshold = 0.3
'succ': 223, 'fail': 218, 'wrong': 1, 'ratio': 0.5056689342403629, 'avg_diff': 0.03405453556221429

targeted attack; pernultimate layer firing thresh = 1.0
'succ': 413, 'fail': 37,  'wrong': 0, 'ratio': 0.9177777777777778, 'avg_diff': 0.033113396922735266

Paper link

https://ieeexplore.ieee.org/document/9527394

DOI 10.1109/TNNLS.2021.3106961

liangling76 / snn_attack