Disruptor

This is an Adversarial Machine Learning tool that can be used to attack a training set used to build a Machine Learning model. Thus, this application can be used to study different attacks and find countermeasures to them.

To do so, this code already contains some attacks, but the tool is designed to be extended with personal attack implementations.

Use it from CLI

Download from the release section the latest disruptor.jar and run the following command for a simple execution:

java -jar disruptor.jar datsetPath.csv

To add more configurations, open the man page of the application with this command:

java -jar disruptor.jar -v

Use it as a library

Maven

Add github as a Maven repository in your pom.xml:

<repositories>
    <repository>
        <id>github</id>
        <url>https://maven.pkg.github.com/gabrielezarcone/disruptor</url>
        <snapshots>
            <enabled>true</enabled>
        </snapshots>
    </repository>
</repositories>

Then add the the following dependency:

<dependencies>
    <dependency>
        <groupId>gzarcone</groupId>
        <artifactId>disruptor</artifactId>
        <version>1.0.0</version>
    </dependency>
</dependencies>

Extend the attacks

Extend the Disruptor class to change the attack list adding the new attack class named NewAttack:

public class NewDisruptor extends Disruptor {
    @Override
    protected void populateAttacksList(Instances dataset, double[][] selectedFeatures) {
        
        super.populateAttacksList(dataset, selectedFeatures);
        NewAttack newAttack = new NewAttack(dataset);
        newAttack.setFeatureSelected(selectedFeatures);
        getAttacksList().add(newAttack);

    }
}

The NewAttack class should be an extension of the abstract class Attack:

public class NewAttack extends Attack {
    protected NewAttack(Instances target) {
        super(target);
    }

    protected NewAttack(Instances target, double capacity, double featuresCapacity, double knowledge) {
        super(target, capacity, featuresCapacity, knowledge);
    }

    @Override
    public Instances attack() {
        // insert here the attack implementation
        return null;
    }
}

Packages

• attacks: contains classes that can be used to create new attacks and some custom attack implementation
• attributeselection: contains classes used to perform attribute selection
• core: the main package that contains the Disruptor class
• costants: the classes of this package contains constants
• experiment: contains the classes used for the evaluation of the attacks
• filters: wrappers of the Weka filters
• perturbeddataset: business object used to transport perturbed dataset together with attacks metadata
• properties: classes used to generate and fetch properties from config files
• roc: classes used for the ROC curves generation (not working at the moment)
• saver: classes used to save and export files
• util: utility classes