Versatile-Weight-Attack

The implementatin of Versatile Weight Attack via Flipping Limited Bits, including Singel Sample Attack and Triggered Samples Attack.

Updates

[2023/05/01] The attacks against full-precision netoworks (their weights are in 32-bit floating point format) have been added to the folders "single_sample_attack_float" and "triggered_samples_attack_float".

Install

1. Install PyTorch >= 1.5
2. Clone this repo:

git clone https://github.com/jiawangbai/Versatile-Weight-Attack.git

Run Our Code

Set the "cifar_root" in the "config.py" firstly.

Single Sample Attack (SSA)

Running the below command will attack a sample (3676-th sample in the CIFAR-10 validation set) into class 0.

cd single_sample_attack/
python SSA.py --target-class 0 --attack-idx 3676 --lam 100 --k 5

You can set "target-class" and "attack-idx" to perform SSA on a specific sample.

Running the below command can reproduce our results in attacking the 8-bit quantized ResNet on CIFAR-10 with the parameter searching strategy introduced in the paper.

cd single_sample_attack/
python SSA_reproduce.py 

"cifar_attack_info.txt" includes the 1,000 attacked samples and their target classes used in our experiments.
Format:
  [[target-class, sample-index],
  [target-class, sample-index],
  ...
  [target-class, sample-index] ]
where "sample-index" is the index of this attacked sample in CIFAR-10 validation set.

Running the below command will attack the full-precision network.

cd single_sample_attack_float/
python SSA.py --target-class 0 --attack-idx 3676 --lam 50 --k 5

Triggered Samples Attack (TSA)

Running the below command will attack all samples with a trigger (CIFAR-10 validation set) into class 0.

cd triggered_samples_attack/
python TSA.py --target 0

You can set "target" to perform TSA with different target class.

Running the below command will reproduce our results in attacking the 8-bit quantized ResNet on CIFAR-10 with the parameter searching strategy introduced in the paper.

cd triggered_samples_attack/
sh run.sh [gpu-id]

Running the below command will attack the full-precision network.

cd triggered_samples_attack_float/
python TSA.py --target 0

Others

We provide the pretrained 8-bit quantized ResNet on CIFAR-10. -> "cifar_resnet_quan_8/model.th" and the pretrained full-precision ResNet on CIFAR-10. -> "cifar_resnet_float/model.th"

Python version is 3.6.10 and the main requirments are below:
  torch==1.5.0
  bitstring==3.1.7
  torchvision==0.6.0a0+82fd1c8
  numpy==1.18.1

We also provide the following command to install dependencies before running the code:

pip install -r requirements.txt