A simple example of running our proposed method is shown as follows.

python FEAWAD.py --network_depth=4 --runs=10 --known_outliers=30 --cont_rate=0.02 --data_format=0 --output=./results.csv --data_set nslkdd_normalization

The meaning of the parameters are shown as follows:

• network_depth: the depth of the network architecture, 1, 2 and 4 available, 4 default.
• batch_size: batch size used in SGD, 512 default.
• nb_batch: the number of batches per epoch, 20 default.
• epochs: the number of epochs (in the end-to-end training stage), 30 default.
• runs: how many times we repeat the experiments to obtain the average performance, 10 default.
• known_outliers: the number of labeled outliers available at hand, 30 default.
• cont_rate: the outlier contamination rate in the training data, 0.02 default.
• input_path: the path of the data sets, './dataset/' default.
• data_set: file name of the dataset chosen, 'nslkdd_normalization' default.
• data_format: specify whether the input data is a csv (0) or libsvm (1) data format, '0' and '1' available, '0' default.
• output: the output file path, './proposed_devnet_auc_performance.csv' default.
• ramdn_seed: the random seed number, 42 default.

The key packages and their versions used in our algorithm implementation are listed as follows

• python==3.6.12
• keras==2.3.1
• tensorflow-gpu==1.13.1
• scikit-learn==0.20.0
• numpy==1.19.4
• pandas==1.1.5
• scipy==1.5.2

See the full paper for the implemenation details of our proposed method.

The full paper can be found in IEEE Xplore or arXiv

The datasets used in our paper are available at the "dataset" folder.

Yingjie Zhou, Xucheng Song, Yanru Zhang, Fanxing Liu, Ce Zhu and Lingqiao Liu. Feature Encoding with AutoEncoders for Weakly-supervised Anomaly Detection, IEEE Transactions on Neural Networks and Learning Systems, 2021.

If you have any question, please email to Prof. Yingjie Zhou (email: [email protected]) or Mr. Fanxing Liu (email: [email protected]).