Readme.md

Real-Time Evasion Attacks against Deep Learning-Based Anomaly Detection from Distributed System Logs

This repository contains LAM (Log Anomaly Mask), a real-time evasion attack against deep learning based anomaly detection from distributed system logs. Please refer to the paper for information on algorithms. The code here contains a LAM evasion attack against DeepLog model for Hadoop File System (HDFS) log dataset (i.e., in data folder). The main code of our attack is in LAM.py. The requirements for the experiments are in requirements.txt.

Running the code

The code can be run by executing run_LAM_experiments.sh. The overall experiments have four steps.

1. Train the deep learning based anomaly detection model (e.g., DeepLog)
2. Obtain anomaly detection results before attack.
3. Train the Perturber for LAM attack.
4. Obtain anomaly detection results after attack.

Note that the current code contains pre-trained anomaly detection models and LAM attack models which are stored in folders models and rl_model respectively. You can turn logging on\off within the code. The logs for the model training will be stored in logs and rl_data for the anomaly detection model and the LAM attack respectively.

Results

The results folder contains the following.

• A trace for training LAM attack.
• Anomaly detection results before and after LAM attack.
• The logs modified after attack stored in adversarially_modified_logs.txt

@inproceedings{herath2021real,
  title={Real-Time Evasion Attacks against Deep Learning-Based Anomaly Detection from Distributed System Logs},
  author={Herath, J Dinal and Yang, Ping and Yan, Guanhua},
  booktitle={Proceedings of the Eleventh ACM Conference on Data and Application Security and Privacy},
  pages={29--40},
  year={2021}
}