============= EvadeML v1.0

---

An Evolutionary Framework for Evading Machine Learning-based Malware Classifiers

Weilin Xu, Yanjun Qi, and David Evans
University of Virginia

Installation

Several external libraries are required in the project.

• A modified version of pdfrw for parsing PDF at https://github.com/mzweilin/pdfrw
• Cuckoo Sandbox v1.2 as the oracle at https://github.com/cuckoosandbox/cuckoo/releases/tag/1.2
• Target classifier PDFrate-Mimicus at https://github.com/srndic/mimicus
• Target classifier Hidost at https://github.com/srndic/hidost

Configuration

Copy the template and change to your own configuration.

cp project.conf.template project.conf
vim project.conf

Running

First start the centralized detection agent with pre-defined malware signatures.

$ ./utils/detection_agent_server.py ./utils/36vms_sigs.pickle

Second, run a program to select several benign PDF files as external genome.

$ ./utils/generate_ext_genome.py [classifier_name] [benign_sample_folder] [file_number]

Now we can start the main program ./gp.py with a long list of arguments. The helper script ./batch.py should be helpful in large scale experiments.

./batch.py [classifier_name] [ext_genome_folder] [round_id]

Add a new classifier to evade

Adding more target classifiers to the framework is trivial.

1. Add a wrapper in ./classifiers/ like pdfrate_wrapper.py::pdfrate()
2. Implement a fitness function in ./lib/fitness.py like fitness_pdfrate(), and specify a switch in gp.py
3. Import the wrapper function in ./utils/detection_agent_server.py like pdfrate(), and extend query_classifier() so that the main program could call the detector through lib.detector.query_classifier().