lironber/goad

Official implementation of "Classification-Based Anomaly Detection for General Data" by Liron Bergman and Yedid Hoshen, ICLR 2020.

GOAD

This repository contains a PyTorch implementation of the method presented in "Classification-Based Anomaly Detection for General Data" by Liron Bergman and Yedid Hoshen, ICLR 2020.

Requirements

• Python 3 +
• Pytorch 1.0 +
• Tensorflow 1.8.0 +
• Keras 2.2.0 +
• sklearn 0.19.1 +

Training

To replicate the results of the paper on the tabular-data:

python train_ad_tabular.py --n_rots=64 --n_epoch=25 --d_out=64 --ndf=32 --dataset=kdd 
python train_ad_tabular.py --n_rots=256 --n_epoch=25 --d_out=128 --ndf=128 --dataset=kddrev
python train_ad_tabular.py --n_rots=256 --n_epoch=1 --d_out=32 --ndf=8 --dataset=thyroid
python train_ad_tabular.py --n_rots=256 --n_epoch=1 --d_out=32 --ndf=8 --dataset=arrhythmia 

To replicate the results of the paper on CIFAR10:

python train_ad.py --m=0.1

Citation

If you find this useful, please cite our paper:

@inproceedings{bergman2020goad,
  author    = {Liron Bergman and Yedid Hoshen},
  title     = {Classification-Based Anomaly Detection for General Data},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2020}
}

goad's Issues

A possible problem with computing the means for evalutaion

During the training you perform the transformation, and then sum the means per transformation within the batch:

for i in range(0, len(train_xs), self.batch_size):
    ...
    tc_zs, ce_zs = self.netC(xs)
    sum_zs = sum_zs + tc_zs.mean(0)
    ...
means = sum_zs.t() / n_batch

However, the self.netC changes between batches, so as far as I can tell, means consists of sums over the outputs of different networks.

Am I wrong?
Thanks!

About tc_loss

I get it

Inconsistency on contamination ratio between paper and code.

Hello,

First of all, thanks for releasing the codebase for your paper. I would like to clarify the definition of contamination ratio as I found they are inconsistent between the paper description and the code.

In the paper page 8, it is told that "To evaluate the robustness of our method to this unsupervised scenario, we analysed the KDDCUP99 dataset, when X% of the training data is anomalous."

In your implementation, however, c_percent of samples from anomaly set is sampled and included to the training data.

n_contaminated = int((c_percent/100)*len(anom_samples))

This seems inconsistent with the description in the paper as normal and anomaly dataset sizes are not equal (e.g., for KDD, there are 4x more normal data than anomaly, so the contamination ratio should be 4x smaller than c_percent if following the description in the paper).

It would be great if this can be clarified. Also, due to this inconsistency, it is difficult to reproduce the contaminated data results on KDDRev or Arrhythmia. It would be great if authors can provide the code that generates the train/val splits for other datasets as well.

Occupy too much memory

I use half the number of transforms(72->36) and I find different classes are affected differently

args for fashion-mnist

Hi, lironber, I reproduce the experiment on Fashion-mnist, but it fails to close to your results in paper, can you share me the setting about Fashion-mnist ? Thank you!