Spectral Adversarial Data Augmentation (SADA)

This repository includes open-source codes, detailed experimental results and full references of our AAAI 2023 work

When Neural Networks Fail to Generalize? A Model Sensitivity Perspective.

News!

SAMix, an variant of SADA is accepted by MICCAI-2023!

Please check SAMix.

Overview

The figure above summarizes our algorithm comparisons framework, Spectral Adversarial Data Augmentation.

• First, our method computes model sensitivity map for the ERM model.
• Then, we run SADA to generate the augmented data and train the model with the JS-div regularization.

Prerequisites

• Python 3.8
• PyTorch 1.8.1+
• A computing device with GPU

Getting started

Installation

• (Not necessary) Install Anaconda3
• Install CUDA 11.0+
• Install PyTorch 1.8.1+

Noted that our code is tested based on PyTorch 1.8.1

Dataset & Preparation

All datasets used in our work are publicly available. In our experiments, we followed the data precrossing in the RandConv.

Measure the sensitivity map of an ERM model

Train and evaluate an ERM model by

python ./AAAIcodeSubmissoin__model_sensitivity_map/train_ERM.py

The model will be saved in ./AAAIcodeSubmissoin__model_sensitivity_map/save_dir.

Measure the sensitivity map of an ERM model by

python ./AAAIcodeSubmissoin__model_sensitivity_map/model_sensitivity_map.py

If you want to try SADA directly, a example of model sensitivity map of an ERM model trained on DIGTS dataset is provided as in

./AAAIcodeSubmissoin__model_sensitivity_map/sample/sample_map.pth

Train model with Spectral Adversarial Data Augmentation (SADA)

Train and evaluate the models with SADA by

python ./AAAIcodeSubmissoin__SADA/train_SADA.py

The key SADA data augmentation module is in

./AAAIcodeSubmissoin__SADA/sada.py

Augmentation settings for all datasets:

• --epsilon iteration of the checkpoint to load. #Default: 0.2
• --step_size step size of the adversarial attack on the amplitude spectrum. #Default: 0.08
• --num_steps batch size of the attack steps. #Default: 5
• --tau settings for the early stop acceleration. #Default: 1
• --random_init if or not initializing amplitude spertrums with random perturbations. #Default: True
• --randominit_type type of random init type. #Default: 'normal'
• --criterion loss functions to attack. #Default: 'ce', choices=['ce', 'kl']
• --direction neg: standard attack || pos:inverse attack. #Default: neg, choices=['pos','neg']

Citation

Please cite these papers in your publications if it helps your research:

@inproceedings{zhang2023neural,
  title={When neural networks fail to generalize? a model sensitivity perspective},
  author={Zhang, Jiajin and Chao, Hanqing and Dhurandhar, Amit and Chen, Pin-Yu and Tajer, Ali and Xu, Yangyang and Yan, Pingkun},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={37},
  number={9},
  pages={11219--11227},
  year={2023}
}

Acknowledgement

We directly adopt some codes from the previous works as below:

• AugMix
• AugMax
• GUD
• M-ADA
• RandConv

We would like to thank these authors for sharing their codes.