Deep One-Class Classification

Update: We now also have implemented Deep SVDD in PyTorch which can be found at https://github.com/lukasruff/Deep-SVDD-PyTorch

This repository provides the implementation of the soft-boundary Deep SVDD and one-class Deep SVDD method we used to perform the experiments for our ”Deep One-Class Classification” ICML 2018 paper. The implementation uses the
Theano and Lasagne libraries.

Citation and Contact

You find the PDF of the Deep One-Class Classification ICML 2018 paper at http://proceedings.mlr.press/v80/ruff18a.html.

If you use our work, please also cite the ICML 2018 paper:

@InProceedings{pmlr-v80-ruff18a,
  title     = {Deep One-Class Classification},
  author    = {Ruff, Lukas and Vandermeulen, Robert A. and G{\"o}rnitz, Nico and Deecke, Lucas and Siddiqui, Shoaib A. and Binder, Alexander and M{\"u}ller, Emmanuel and Kloft, Marius},
  booktitle = {Proceedings of the 35th International Conference on Machine Learning},
  pages     = {4393--4402},
  year      = {2018},
  volume    = {80},
}

If you would like to get in touch, please contact contact@lukasruff.com.

Abstract

Despite the great advances made by deep learning in many machine learning problems, there is a relative dearth of deep learning approaches for anomaly detection. Those approaches which do exist involve networks trained to perform a task other than anomaly detection, namely generative models or compression, which are in turn adapted for use in anomaly detection; they are not trained on an anomaly detection based objective. In this paper we introduce a new anomaly detection method—Deep Support Vector Data Description—, which is trained on an anomaly detection based objective. The adaptation to the deep regime necessitates that our neural network and training procedure satisfy certain properties, which we demonstrate theoretically. We show the effectiveness of our method on MNIST and CIFAR-10 image benchmark datasets as well as on the detection of adversarial examples of GTSRB stop signs.

Installation

This code is written in Python 2.7 and requires the packages listed in requirements.txt in the given versions. We recommend to set up a virtual environment in which the packages are then installed, e.g. using virtualenv:

virtualenv -p python2 env
pip install --upgrade https://github.com/Lasagne/Lasagne/archive/master.zip
pip install requests
pip install -r requirements.txt

We install Lasagne first as the 0.2.dev1 version is only available from the GitHub Lasagne repository.

To acitvate/deactivate the environment, run

source env/bin/activate
source deactivate

Make sure that Theano floats are set to float32 by default. This can be done by adding

[global]
floatX = float32

to the ~/.theanorc configuration file that is located in your home directory (you may have to create the .theanorc-file if you have not used Theano before).

Repository structure

/data

Contains the data. We implemented support for MNIST and CIFAR-10:

To run the experiments, the datasets must be downloaded from the original sources in their original formats into the data folder.

/src

Source directory that contains all Python code and shell scripts to run experiments.

/log

Directory where the results from the experiments are saved.

To reproduce results

Change your working directory to src and make sure that the respective datasets are downloaded into the data directory. The src directory has two subfolders src/experiments and src/scripts. The scripts directory provides interfaces to run the implemented methods on different datasets with different settings.

The interface for our Deep SVDD method is given by:

sh scripts/mnist_svdd.sh ${device} ${xp_dir} ${seed} ${solver} ${lr} ${n_epochs} ${hard_margin} ${block_coordinate} ${pretrain} ${mnist_normal} ${mnist_outlier};

For example to run a MNIST experiment with 0 as the normal class (mnist_normal = 0) and all other classes considered to be anomalous (mnist_outlier = -1), execute the following line:

sh scripts/mnist_svdd.sh cpu mnist_0vsall 0 adam 0.0001 150 1 0 1 0 -1;

This runs a one-class Deep SVDD (hard_margin = 1 and block_coordinate = 0) experiment with pre-training routine (pretrain = 1) where one-class Deep SVDD is trained for n_epochs = 150 with the Adam optimizer (solver = adam) and a learning rate of lr = 0.0001. The experiment is executed on device = cpu and results are exported to log/mnist_0vsall. For reproducibility, we set seed = 0.

You find descriptions of the various script options within the respective Python files that are called by the shell scripts (e.g. baseline.py).

The experiments directory can be used to hold shell scripts that start multiple experiments at once. For example 

sh experiments/mnist_svdd_exp.sh

starts all MNIST one-class Deep SVDD experiments (i.e. ten one-class classification setups each for ten seeds). This particular script runs the experiments on 10 CPUs in parallel.

Disclosure

This implementation is based on the repository https://github.com/oval-group/pl-cnn, which is licensed under the MIT license. The pl-cnn repository is an implementation of the paper Trusting SVM for Piecewise Linear CNNs by Leonard Berrada, Andrew Zisserman and M. Pawan Kumar, which was an initial inspiration for this research project.