Skip to content

Latest commit

 

History

History
187 lines (143 loc) · 6.98 KB

how_to.md

File metadata and controls

187 lines (143 loc) · 6.98 KB
Raw

How to run the code

Requirements

Install SpikingJelly repo from here, or install it using pip (included in the requirements):

Install the requirements:

pip install -r requirements.txt

Preparing the datasets

Some datasets are automatically downloaded. But some others have to be downloaded manually. This is a restriction from the SpikingJelly repo.

Create a data/ folder in the root of the project:

mkdir data

Then, the datasets should be placed in the following format: data/<dataset>/download/

Check the following section for more details about specific datasets.

N-MNIST

N-MNIST cannot be downloaded automatically. You have to download it manually from here. Then, create a folder with the name mnist in the data folder and put the downloaded files in it.

mkdir data/mnist

And put the dataset (.zip file) in it. SpikingJelly will automatically unzip the files (creating a extracted/ folder), and do the rest of the work.

CIFAR10-DVS

CIFAR10-DVS can be downloaded automatically. You just have to create a folder with the name cifar10 in the data folder.

mkdir data/cifar10

DVS128 Gesture

DVS128 Gesture cannot be downloaded automatically. You have to download it from here. Then, create a folder with the name gesture in the data folder and put the downloaded files in it.

mkdir data/gesture

And put the dataset (.gz file) in it.

N-Caltech101

N-Caltech101 cannot be downloaded automatically. You have to download it from here. Then, create a folder with the name caltech in the data folder and put the downloaded files in it.

mkdir data/caltech

And put the dataset (.zip file) in it.

Hardware requirements

In order to run the code a GPU is strongly recommended. The code is tested on a machine with 1 NVIDIA A100 GPUs with 40GB.

Reading the results

After execution, the results are saved in a .csv file in the experiments/ folder. The .csv file will contain all the execution parameters and the results of the attack, i.e., clean accuracy and backdoor accuracy.

Examples

Script examples for different datasets are provided in the scripts/ folder.

Get help:

python main.py --help

usage: main.py [-h] [--dataset DATASET] [--lr LR] [--batch_size BATCH_SIZE] [--epochs EPOCHS] [--T T]
               [--amp] [--cupy] [--loss {mse,cross}] [--optim {adam,sgd}] [--trigger_label TRIGGER_LABEL]
               [--polarity {0,1,2,3}] [--trigger_size TRIGGER_SIZE] [--epsilon EPSILON]
               [--pos {top-left,top-right,bottom-left,bottom-right,middle,random}] [--type {static,moving,smart}]
               [--n_masks N_MASKS] [--least] [--most_polarity] [--momentum MOMENTUM] [--data_dir DATA_DIR]
               [--save_path SAVE_PATH] [--model_path MODEL_PATH] [--seed SEED]

optional arguments:
  -h, --help            show this help message and exit
  --dataset DATASET     Dataset to use
  --lr LR               Learning rate
  --batch_size BATCH_SIZE
                        Batch size
  --epochs EPOCHS       Number of epochs
  --T T                 simulating time-steps
  --amp                 Use automatic mixed precision training
  --cupy                Use cupy
  --loss {mse,cross}    Loss function
  --optim {adam,sgd}    Optimizer
  --trigger_label TRIGGER_LABEL
                        The index of the trigger label
  --polarity {0,1,2,3}  The polarity of the trigger
  --trigger_size TRIGGER_SIZE
                        The size of the trigger as the percentage of the image size
  --epsilon EPSILON     The percentage of poisoned data
  --pos {top-left,top-right,bottom-left,bottom-right,middle,random}
                        The position of the trigger
  --type {static,moving,smart}
                        The type of the trigger
  --n_masks N_MASKS     The number of masks. Only if the trigger type is smart
  --least               Use least active area for smart attack
  --most_polarity       Use most active polarity in the area for smart attack
  --momentum MOMENTUM   Momentum
  --data_dir DATA_DIR   Data directory
  --save_path SAVE_PATH
                        Path to save the experiments
  --model_path MODEL_PATH
                        Use a pretrained model
  --seed SEED           Random seed

Static Triggers

Example of running a static trigger attack on N-MNIST dataset, in the top-left corner, with 10% of the data poisoned, polarity 1, with the trigger size of 10% of the image size:

python main.py --dataset mnist --polarity 1 --pos top-left --trigger_size 0.1 --epsilon 0.1 --type static --cupy --epochs 10

Moving Triggers

Example of running a moving trigger attack on N-MNIST dataset, in the top-left corner, with 10% of the data poisoned, polarity 1, with the trigger size of 10% of the image size:

python main.py --dataset mnist --polarity 1 --pos top-left --trigger_size 0.1 --epsilon 0.1 --type moving --cupy --epochs 10

Smart Triggers

Example of running a smart trigger attack on N-MNIST dataset, in the the least important area, with 10% of the data poisoned, with the trigger size of 10% of the image size:

python main.py --dataset mnist --trigger_size 0.1 --epsilon 0.1 --type smart --least --cupy --epochs 10

Dynamic Triggers

Get help:

usage: dynamic.py [-h] [--dataset DATASET] [--lr LR] [--batch_size BATCH_SIZE] [--epochs EPOCHS]
                  [--train_epochs TRAIN_EPOCHS] [--T T] [--amp] [--cupy] [--loss {mse,cross}] [--optim {adam,sgd}]
                  [--trigger_label TRIGGER_LABEL] [--momentum MOMENTUM] [--alpha ALPHA] [--beta BETA]
                  [--data_dir DATA_DIR] [--save_path SAVE_PATH] [--seed SEED]

optional arguments:
  -h, --help            show this help message and exit
  --dataset DATASET     Dataset to use
  --lr LR               Learning rate
  --batch_size BATCH_SIZE
                        Batch size
  --epochs EPOCHS       Number of epochs
  --train_epochs TRAIN_EPOCHS
                        Number of epochs
  --T T                 simulating time-steps
  --amp                 Use automatic mixed precision training
  --cupy                Use cupy
  --loss {mse,cross}    Loss function
  --optim {adam,sgd}    Optimizer
  --trigger_label TRIGGER_LABEL
                        The index of the trigger label
  --momentum MOMENTUM   Momentum
  --alpha ALPHA         Alpha
  --beta BETA           Beta. Gamma in the paper
  --data_dir DATA_DIR   Data directory
  --save_path SAVE_PATH
                        Path to save the experiments
  --seed SEED           Random seed

Example of running a dynamic trigger attack on N-MNIST dataset:

python dynamic.py --dataset mnist --cupy --epochs 10 --train_epochs 1 --alpha 0.5 --beta 0.01