MemStream

Implementation of

• MemStream: Memory-Based Streaming Anomaly Detection. Siddharth Bhatia, Arjit Jain, Shivin Srivastava, Kenji Kawaguchi, Bryan Hooi. The Web Conference (formerly WWW), 2022.

MemStream detects anomalies from a multi-aspect data stream. We output an anomaly score for each record. MemStream is a memory augmented feature extractor, allows for quick retraining, gives a theoretical bound on the memory size for effective drift handling, is robust to memory poisoning, and outperforms 11 state-of-the-art streaming anomaly detection baselines.

After an initial training of the feature extractor on a small subset of normal data, MemStream processes records in two steps: (i) It outputs anomaly scores for each record by querying the memory for K-nearest neighbours to the record encoding and calculating a discounted distance and (ii) It updates the memory, in a FIFO manner, if the anomaly score is within an update threshold β.

Demo

1. KDDCUP99: Run python3 code/memstream.py --dataset KDD --beta 1 --memlen 256
2. NSL-KDD: Run python3 code/memstream.py --dataset NSL --beta 0.1 --memlen 2048
3. UNSW-NB 15: Run python3 code/memstream.py --dataset UNSW --beta 0.1 --memlen 2048
4. CICIDS-DoS: Run python3 code/memstream.py --dataset DOS --beta 0.1 --memlen 2048
5. SYN: Run python3 code/memstream-syn.py --dataset SYN --beta 1 --memlen 16
6. Ionosphere: Run python3 code/memstream.py --dataset ionosphere --beta 0.001 --memlen 4
7. Cardiotocography: Run python3 code/memstream.py --dataset cardio --beta 1 --memlen 64
8. Statlog Landsat Satellite: Run python3 code/memstream.py --dataset statlog --beta 0.01 --memlen 32
9. Satimage-2: Run python3 code/memstream.py --dataset satimage-2 --beta 10 --memlen 256
10. Mammography: Run python3 code/memstream.py --dataset mammography --beta 0.1 --memlen 128
11. Pima Indians Diabetes: Run python3 code/memstream.py --dataset pima --beta 0.001 --memlen 64
12. Covertype: Run python3 code/memstream.py --dataset cover --beta 0.0001 --memlen 2048

Command line options

• --dataset: The dataset to be used for training. Choices 'NSL', 'KDD', 'UNSW', 'DOS'. (default 'NSL')
• --beta: The threshold beta to be used. (default: 0.1)
• --memlen: The size of the Memory Module (default: 2048)
• --dev: Pytorch device to be used for training like "cpu", "cuda:0" etc. (default: 'cuda:0')
• --lr: Learning rate (default: 0.01)
• --epochs: Number of epochs (default: 5000)

Input file format

MemStream expects the input multi-aspect record stream to be stored in a contains , separated file.

Datasets

Processed Datasets can be downloaded from here. Please unzip and place the files in the data folder of the repository.

1. KDDCUP99
2. NSL-KDD
3. UNSW-NB 15
4. CICIDS-DoS
5. Synthetic Dataset (Introduced in paper)
6. Ionosphere
7. Cardiotocography
8. Statlog Landsat Satellite
9. Satimage-2
10. Mammography
11. Pima Indians Diabetes
12. Covertype

Environment

This code has been tested on Debian GNU/Linux 9 with a 12GB Nvidia GeForce RTX 2080 Ti GPU, CUDA Version 10.2 and PyTorch 1.5.