ELISE-2021

This repository is the source code for ELISE.

Datasets

We provide part of our datasets in Driver. Because log files contain a lot of sensitive information, we removed them for privacy concerns.

Deployment

We use a similar environment as DeepZip and DeepZip provides a docker environment (We copy the makefile file and docker file from DeepZip).

Install docker

• For bash environment

cd docker
make bash

• For other environments, please refer to the implementation of DeepZip.
• For reuse of container.

docker start containerID
docker exec -it containerID /bin/bash

Example of compression.

• Audit log file.

cd audit
#preprocessing
python extract_audit.py -input audit_log_1
python parse_audit.py -input audit_log_1 -output audit_tmp.npy -param audit_tmp.params.json
#training
cd ../
python trainer_elise.py -gpu 0 -param audit/audit_tmp.params.json -d audit/audit_tmp.npy -name audit/audit_tmp.hdf5 -model_name LSTM_multi_bn -batchsize 4096 -log_file LSTM.log.csv
#compression
cd audit
./compress_audit.sh
#decompression
./decompress_audit.sh
#check correctness
python compare_audit.py

• Different log files.

#preprocessing
nohup python -u parse_ftp.py -input ../../ftp_1 -param ftp_tmp.params.json -output ftp_tmp.npy &

nohup python -u parse_darpa.py -input ../../darpa1 -number 1 -param darpa_tmp.params.json -output darpa_tmp.npy &

#training 
nohup python -u trainer_elise.py -gpu 0 -file_type ftp -param ./ftp/ftp_tmp.params.json -d ./ftp/ftp_tmp.npy -name ./ftp/ftp_tmp.hdf5 -model_name LSTM_multi_bn -batchsize 4096 -log_file LSTM.log.csv &

nohup python -u trainer_elise.py -gpu 0 -param ./bsd/darpa_tmp.params.json -d ./bsd/darpa_tmp.npy -name ./bsd/darpa_tmp.hdf5 -model_name LSTM_multi_bn -batchsize 4096 -log_file LSTM.log.csv &

nohup python -u trainer_elise.py -file_type httpd -gpu 0 -epoch 4 -d ./http/httpd_tmp.npy -param ./http/httpd_tmp.params.json -name ./http/httpd_tmp.hdf5 -model_name LSTM_multi_bn -batchsize 4096 -log_file LSTM.log.csv &

nohup python -u trainer_elise.py -gpu 0 -file_type windows -param ./win/win_tmp.params.json -d ./win/win_tmp.npy -name ./win/win_tmp.hdf5 -model_name LSTM_multi_bn -batchsize 4096 -log_file LSTM.log.csv &

nohup python -u trainer_elise.py -gpu 3 -param ./sql/mysql_tmp.params.json -d ./sql/mysql_tmp.npy -file_type mysql -name ./sql/sql_tmp.hdf5 -model_name LSTM_multi_bn -batchsize 4096 -log_file LSTM.log.csv &


#compression
nohup python -u compressor.py -data ftp_tmp.npy -gpu 0 -data_params ftp_tmp.params.json -model ftp_tmp.hdf5 -model_name LSTM_multi_bn -output ftp_tmp.compressed -batch_size 1000 &

nohup python -u compressor_bsd.py -data darpa_tmp.npy -gpu 2 -data_params darpa_tmp.params.json -model darpa_tmp.hdf5 -model_name LSTM_multi_bn -output darpa_tmp.compressed -batch_size 1000 &

nohup python -u compressor.py -data httpd_tmp.npy -gpu 5 -data_params httpd_tmp.params.json -model httpd_tmp.hdf5 -model_name LSTM_multi_bn -output httpd_tmp.compressed -batch_size 1000 &

nohup python -u compressor.py -data mysql_tmp.npy -gpu 0 -data_params mysql_tmp.params.json -model sql_tmp.hdf5 -model_name LSTM_multi_bn -output sql_tmp.compressed -batch_size 1000 &

nohup python -u compressor.py -data win_tmp.npy -gpu 5 -data_params win_tmp.params.json -model win_tmp.hdf5 -model_name LSTM_multi_bn -output win_tmp.compressed -batch_size 1000 &

#decompression
nohup python -u decompressor_ftp.py -output ftp_tmp_decompress -gpu 0 -model ftp_tmp.hdf5 -model_name LSTM_multi_bn -input_file_prefix ftp_tmp.compressed -batch_size 1000 &

nohup python -u decompressor_darpa.py -gpu 2 -number 1 -output darpa_tmp_decompress -model darpa_tmp.hdf5 -model_name LSTM_multi_bn -input_file_prefix darpa_tmp.compressed -batch_size 1000 &

nohup python -u decompressor_httpd.py -output httpd_tmp_decompress -gpu 5 -model httpd_tmp.hdf5 -model_name LSTM_multi_bn -input_file_prefix httpd_tmp.compressed -batch_size 1000 & 

nohup python -u decompressor_mysql.py -output sql_tmp_decompress -gpu 0 -model sql_tmp.hdf5 -model_name LSTM_multi_bn -input_file_prefix sql_tmp.compressed -batch_size 1000 &

nohup python -u decompressor_windows.py -gpu 3 -output win_tmp_decompress -model win_tmp.hdf5 -model_name LSTM_multi_bn -input_file_prefix win_tmp.compressed -batch_size 1000 &

Note that key patterns can be automatically found when parsing logs into JSON format. To illustrate, we provide a python file to extract key patterns of our datasets to avoid users additionally using a parsing tool such as logstash. We also provide processed key pattern files so that users do not need to run the code to obtain key patterns again.

Some important settings

For preprocessing 4, matching of substrings is complicated. Considering the efficiency, we implement a simpler way to get frequent/common long strings directly.

Also, we provide several comparison files to help users compare the contents before and after compression. Since preprocessing 2 sorts the log entries and divides them into different sessions, to keep the exact same sequence of log entries, users can either sort log entries with their sequence ids/timestamps or create additional indexes during preprocessing.

Other resources

The implementation of DeepZip can be found at DeepZip. We also provide a more memory efficient training file (trainer.py).