Learning from Convolution-Based Unlearnable Datasets

This repository contains the implementation of the paper Learning from Convolution-Based Unlearnable Datasets.

It was accepted to the NeurIPS Workshop, The 3rd New Frontiers in Adversarial Machine Learning 2024 (AdvML Frontiers @ NeurIPS 2024).

Dohyun Kim, Pedro Sandoval-Segura

Use RSK and SSK + FF

If you want to use our SSK + FF or RSK + FF transforms, use the following code:

RSK_FF = transforms.Compose([ToTensor(), 
         transforms.Lambda(lambda x: sharpen_image(x, center_mean=2.5, random=True)),
         transforms.Lambda(lambda x: get_dct_image(x, ff_percentage=30))])
SSK_FF = transforms.Compose([ToTensor(), 
         transforms.Lambda(lambda x: sharpen_image(x, center_mean=2.5, random=False)), 
         transforms.Lambda(lambda x: get_dct_image(x, ff_percentage=30))])

and copy the sharpen_image and get_dct_image functions from augments.py.

Usage

You can customize the training process using various command-line arguments. Here are the options:

• --epochs: Number of training epochs (default: 60)
• --dataset: Choose between 'cifar10' and 'cifar100' (default: 'cifar10')
• --normalize: Normalize the dataset (after transforms) (default: True)
• --ud: Choose between 'cuda' (for CUDA dataset) and 'clean' (default: 'cuda')
• --blur_parameter: Blur parameter for CUDA dataset (default: 0.3)
• --sharp-center: Changes the center value for the sharpening kernel (default: 2.5)
• --ff_percentage: Sets the percentage of lowest frequencies kept (default: 30)
• --transform: Augmentation technique to use (choices: 'ssk', 'ssk_ff', 'rsk', 'rsk_ff', 'ff', 'clean', default: 'rsk_ff')

Example:

python main.py --epochs 100 --dataset cifar100 --transform rsk --ud clean

This command will train a ResNet-18 for 100 epochs on the CIFAR-100 dataset using RSK augmentation without CUDA dataset poisoning.

To train the model with default settings on CIFAR-10 dataset:

python main.py

This will train a ResNet-18 on the CUDA dataset with the RSK + FF transform applied.

File Descriptions

• main.py: The main script for training the model.
• augments.py: Contains functions for data augmentation, including RSK and FF.
• resnet.py: Implements ResNet-18
• util.py: Utility functions
• requirements.txt: Lists the Python dependencies for the project.

Installation

1. Create a conda environment:

   conda create -n rsk python=3.10
   conda activate rsk
   

2. Clone this repository:

   git clone https://github.com/aseriesof-tubes/rsk.git
   cd rsk
   

3. Install the required packages:

   pip install -r requirements.txt