Author: Wooheon Hong, Minsoo Kim with Samsung Electronics

Date: 2021.05.10 ~ 2022. 04. 29

This is the pytorch implementation of the PUMAD but, the detailed code cannot be disclosed.

PUMAD: PU Metric learning for anomaly detection(Information Sciences, 2020)

Distance Hashing-based Filtering

• Find reliable normal and potential anomaly in unlabeled data

Deep Metric Learning

• push negative samples and pull positive samples for each anchor(normal, anomaly)

• python 3.8
• see requirements.txt

conda create –n env_name python=3.8
conda activate env_name
pip install --upgrade pip
pip install -r requirements.txt

X: input data [ndarray] *shape (data, # of features, window length) 
y: input labels [ndarray] *shape : (data) 

Measure the AUC at each epochs

python train.py --device <gpu number> --dataset <name of .npy file> --data_path <location of files> --window <length of sliding window> 

The detailed descriptions about the parameters are as following: Note : There is

Evaluation metrics are precision, recall and F1-score.

python test.py save/[CHECKPOINT].pth
python heatmap.py --savedir save --dataset unsw-nb15 --window_size 10 --plotting

Run test.py first, then run heatmap.py.
The parameters of the test.py are same as train.py, except the checkpoint path.
The parameters of the heatmap.py are:
heatmap.py [-h] [--save_dir SAVE_DIR] [--dataset DATASET] [--window_size WINDOW_SIZE] [--_class {,_normal,_anomaly}] [--plotting] \

├── heatmap.py
├── model.py
├── puamd_structure.PNG
├── README.md
├── requirements.txt
├── save
├── test.py
├── train.py
└── utils
    ├── utils_dataset.py
    ├── utils_dhf.py
    └── utils.py

Embedding Model : WaveNet

A file with the main function has an argument setting and a training function.

Make dataset for PUMAD

• Parameters

  • X_data : [ndarray] *shape (data, # of features, window length)
  • y_data : [ndarray] *shape : (data)
  • n_htables : # of the hash tables in DHF
  • n_bits : # of bits in hash
  • transform : embedding model

• Outputs

  • (anchor, pos, neg) : tuple of anchor tensor, positive tensor, negative tensor

Pytorch Dataset class for test dataset

Function for creating data with diverse experiment settings

• Parameters

  • X_train: input train data [ndarray]
  • y_train: input train labels [ndarray]
  • X_test: input test data [ndarray]
  • y_test: input test labels [ndarray]
  • n_labels: # of partially observed anomalies [int]

• Outputs

  • preprocessed_data : [list] *[X_train, y_train, y_train_pu, X_test, y_test]

It is a function of finding reliable normal and potentail abnormal through Distance Hashing based Filtering(DHF) with embedded data.

• Parameters

  • embedded_anom : embedding anomaly data
  • embedded_noise : embedding unlabeled data
  • n_htables : # of the hash tables in DHF
  • n_bits : # of bits in hash

• Outputs

  • reliable normal index
  • potential abnormal index
  • avg_n_buckets : average for the # of hash table buckets