petukhovv/anomaly-detection

Anomaly detection using autoencoder and dbscan / 3(5)-sigma (adapted for Kotlin code anomaly detection)

anomaly-detection

Anomaly detection using autoencoder and dbscan (adapted for Kotlin source code anomaly detection)

Available steps (stages):

• autoencoding: run the autoencoder on the specified dataset, calulating and write vectors with differences (between input and output) or simple euclidean distances array;
• anomaly_selection: read difference vectors or distances vector and anomaly selection in it (via DBScan or 5-sigma);
• [without specify stage]: run of both stages without intermediate write of difference vectors or distances vector in a file.

Program use

Autoencoding

Description

At this stage, the vectors representing the AST are encoded and decoded, and the differences between the input vectors and decoded vectors is written to a file

Stage arguments

• -s, --stage -> autoencoding;
• --use_dbscan (default=false): whether to use dbscan (high memory or time usage!) - then will use full differences between autoencoder input and output vectors matrix; if not, then will use simple euclidean distance between autoencoder input and output vectors;
• -f, --dataset: path to dataset file (csv format with colon delimiter);
• --split_percent: dataset train/test split percent;
• --encoding_dim_percent: encoding dim percent (towards features number);
• --differences_output_file: path to file with input-decoded difference (full differences matrix if --use_dbscan=True or simple distances vector if not).

Example of use

With DBScan:

python3 main.py -s autoencoding --use_dbscan -f dataset.csv --split_percent 0.9 --encoding_dim_percent 0.8 --differences_output_file differences.bin

Without DBScan:

python3 main.py -s autoencoding -f dataset.csv --split_percent 0.9 --encoding_dim_percent 0.8 --differences_output_file distances.json

If use full differences matrix (with --use_dbscan option), then the file will be written in binary mode.

Anomaly selection

Description

At this stage, anomalies are selected by the difference matrix (via DBScan) or the distances vector (via 5-sigma)

Stage arguments

• -s, --stage -> autoencoding;
• --use_dbscan (default=false): whether to use dbscan (high memory or time usage!) - then will use full differences between autoencoder input and output vectors matrix; if not, then will use simple euclidean distance between autoencoder input and output vectors;
• --differences_file: path to file with input-decoded difference (full differences matrix or simple distances vector), obtained previous stage (autoencoding);
• --files_map_file: file with map dataset indexes and ast file paths, obtained by ast-set2matrix with stage=vectors2matrix;
• -o, --anomalies_output_file: path to file, which will contain ranking anomaly list (as paths to AST code snippets and ranks);

Example of use

With DBScan:

python3 main.py -s anomaly_selection --use_dbscan --differences_file differences.bin --files_map_file files_map.json --anomalies_output_file anomalies.json

Without DBScan:

python3 main.py -s anomaly_selection --differences_file distances.json --files_map_file files_map.json --anomalies_output_file anomalies.json

If use full differences matrix (with --use_dbscan option), then the file will be read in binary mode.

Without specify stage

If you do not specify a stage, then runs both stages.

Use arguments both stages except --differences_output_file and --differences_file.

Example of use

With DBScan:

python3 main.py -f dataset.csv --files_map_file files_map.json --split_percent 0.9 --encoding_dim_percent 0.8 --anomalies_output_file anomalies.json

Without DBScan:

python3 main.py --use_dbscan -f dataset.csv --files_map_file files_map.json --split_percent 0.9 --encoding_dim_percent 0.8 --anomalies_output_file anomalies.json