aucad / aml-networks

Evaluating adversarial machine learning attacks in network intrusion detection systems.

Evaluating AML Threats in NIDS

This repository contains an implementation of an evaluation system to run experiments. An experiment measures success rate of adversarial machine learning (AML) evasion attacks in network intrusion detection systems (NIDS). The system allows to evaluate classifiers trained on network data sets against adversarial black-box evasion attacks.

Experiment options

Option	Description
Datasets	Aposemat IoT-23 contains IoT network traffic UNSW-NB15 contains traditional network intrusion data
Classfiers	Keras deep neural network (DNN) A tree-based ensemble learner XGBoost (XGB)
Defenses	For DNN, the defense is Adversarial Training For XGB, the defense is RobustTrees
Attacks	Zeroth-Order Optimization (ZOO) HopSkipJump attack (HSJ)

Source code organization

• .github Actions workflow files
• aml Evaluation system implementation source code
• config Experiment configuration files
• data Preprocessed datasets ready for experiments
• result Referential result for comparison
• RobustTrees (submodule) XGBoost enhanced with adversarial defense

Getting Started

The easiest way to run experiments is with Docker. The docker build assumes amd64-compatible host. Otherwise, build from source.

1. Clone repo and build a container image

git clone https://github.com/aucad/aml-networks.git && cd aml-networks
docker build -t aml-networks . 

2. Launch the container

docker run -v $(pwd)/output:/usr/src/aml-networks/output -it --rm aml-networks /bin/bash

Reproduce Paper Experiments

The runtime estimates are for 8-core 32 GB RAM Linux (Ubuntu 20.04) machine, not using docker; actual times may vary.

1. Limited model queries

make query

[24h] This experiment uses the full testing set and repeats experiments with different model query limits. By default, the max query limits are: 2, 5, default (varies by attack).

2. Limited sampled input

make sample

[90 min] Run experiments using limited input size by randomly sampling the testing set. By default, the sample size is 50 and sampling is repeated 3 times. The result is the average of 3 runs.

3. Plot results

make plots

[1 min] Plot results of the two previous experiments. The plot data source is output/ directory.

Run Custom Experiments

There are three execution modes:

experiment - Performs adversarial attack experiments
plot       - Generate tables from captured experiment results
validate   - Check a dataset for network protocol correctness

Custom experiments can be defined by constructing appropriate commands.

python3 -m aml {experiment|plot|validate} [ARGS]

To see available options for experiments, run:

python3 -m aml experiment --help

To see available options for plotting results, run:

python3 -m aml plot --help

To see available options for the validator, run:

python3 -m aml validate --help

Native Execution

These steps explain how to run experiments from source natively on host machine. You should also follow these steps, if you want to prepare a development environment and make code changes.

Step 0: Environment setup

• 🐍 Required Python environment: 3.8 or 3.9

• ⚠️ Submodule This repository has a submodule. Clone it including the submodule:

  git clone --recurse-submodules https://github.com/aucad/aml-networks.git
  

This implementation is not compatible with Apple M1 machines due to underlying dependency (tensorflow-macos); and although it does not prevent most experiments, some issues may surface periodically.

Step 1: Build robust XGBoost

The evaluation uses a modified version of XGBoost classifier, enhanced with adversarial robustness property. This classifier is not installed with the other package dependencies and must be built locally from source, i.e. the submodule RobustTrees. By default, you will need gcc compiler with OpenMP support. To build robust XGBoost, run:

cd RobustTrees
make -j4

If the build causes issues, follow these instructions to build it from source.

Step 2: Install dependencies

Install required Python dependencies.

python3 -m pip install -r requirements-dev.txt

Install XGBoost from the local build location:

python3 -m pip install -e "/path/to/RobustTrees/python-package"

Step 3: (optional) Check installation

Check the xgboost runtime, the version number should be 0.72.

python3 -m pip show xgboost

Run a help command, which should produce a help prompt.

python3 -m aml

You are ready to run experiments and make code changes.