DeepLog

• This is the achievement of core DeepLog for research aim.

• It is the basic thought with feature engineering to analyse raw logs and finally report the potential malicious logs based on a series of processings.

Ongoing:

• dvc experiments
• dvc dags

Feature:

• convert the logs to structured pandas framework
• extract the log keys from raw logs
• analyse the log key exeuction path
• analyse the paramaters in log key
• analyse the time series data generated from window size and time interval by PCA.
• online learning for feedbacks

For the dataset, I have given some examples and you can put your own data into that folder.

pre-preparation:

# in order to match the libraries versions, please run and build the project in virtual environment
virtualenv env
pip3 install -r requirement.txt

Instructions (In Deeplog_demo folder):

1. Source data:

When the data format is in csv, we need translate them into txt files and split them into batches.

python3 csv_txt_trans.py 

You will get notice on inputing the source location and output location.

2. Data analysis:

we use the logparser tool to transform the source txt log files into structured csv files under a folder, the folder is named by the start and end time. (Find the Lenma_demo under the logparser/logparser/demo)

(use Lenma_demo.py with python2) ---> The python3 version is not provided here. You need to set the locations first:

input_dir = '../../Dataset/Linux/Clear/'   # set the location to yours
output_dir = '../../Dataset/Linux/Clear_Separate_Structured_Logs/'    # set the location to yours

Then you can execute the demo file with python 2.x:

python Lenma_demo.py 

In the stage, we calculate the EventTemplate for every log.

3. Variable Selection:

The log_value_vector.py will be used to generate the csv file, which will be used to implement the anomaly detection later.

(and has been integrated into models already in demo)

4. Model detection:

Basiclly, we have two modules for DeepLog

• Whereas, before implementing the modules, we will first see whether there is obvious malicious logs, we will report them first.

• After that, we will first implement execution path anomaly detection with Execution_Path_Anomaly.py

• Finally, we will implement parameter values anomaly detection with Parameter_value_performance_anomaly.py

• As a plus, there is the ML model using PCA in loglizer.

# go to the folder of model
python3 Execution_Path_Anomaly.py
# go to the folder of model
python3 Parameter_Value_Vector.py 

Statement:

• The model is based on off-line work, the online real-time detection is not available.
• The loglizer and logparser are open source tools, author's rights are reserved.
• I enriched the two tools in the project, notice the differences from the original version.

References：

1.Execution Anomaly Detection in Distributed Systems through Unstructured Log Analysis

2.DeepLog: Anomaly Detection and Diagnosis from System Logs

3.Incremental Construction of LSTM Recurrent Neural Network