This Docker container is a CLI environment featuring a toolkit that gathers executable analyzers, packing detectors, packers and unpackers but also many tools for generating and manipulating datasets of packed and not-packed executables of different formats (including PE, ELF and Mach-O) for the sake of evaluating static detection techniques and tools, visualizing executables' layout and automating machine learning pipelines with the support of many algorithms.
See the Black Hat Arsenal presentations for demonstrations:
Here is what you can see when you start up the Docker container.
The various items integrated in the Packing-Box are defined in the very declarative and easy-to-use YAML format through different configuration files. This makes shaping the scope for evaluations and machine learning model training straightforward and practical for researchers.
Building the image:
# docker build -t dhondta/packing-box .
[...]
<<<wait for a while>>>
[...]
Starting it up with the current working directory mounted as /mnt/share
in the container:
# docker run -it -h packing-box -v `pwd`:/mnt/share dhondta/packing-box
┌──[user@packing-box]──[/mnt/share]──────── ────[172.17.0.2]──[12:34:56]──[0.12]────
$
Items are configured through the YAML configuration files. They consist in:
analyzers.yml
: utilities for analyzing files or more specifically packer tracedetectors.yml
: tools for analyzing and deciding whether an executable is packed or notpackers.yml
and unpackers.yml
: self-explanatoryFrom within the Packing-Box, the packing-box
tool allows to setup and test items.
Operation | Description | Command |
---|---|---|
setup |
Setup an item from its YAML install definition |
# packing-box setup detector die |
test |
Test an item using a built-in set of test samples | # packing-box test packer upx |
Afterwards, items are available from the console.
$ die --help
<<snipped>>
$ upx --help
<<snipped>>
Packers and detectors have their respective dedicated tools for mass operations, packer
and detector
. They work either on a single file, a complete folder or a special dataset instance (as of the abstraction defined in the pbox
package).
$ packer upx path/to/executables --prefix "upx_"
<<snipped>>
For the detector
tool, not selecting any detector will use those selected in detectors.yml
as being part of the "superdetector". Moreover, the --binary
option will consider whether the target executable is packed or not and not is precise packer.
$ detector path/to/single-executable -d die -d pypackerdetect
<<snipped>>
$ detector path/to/executables
<<snipped ; will use "superdetection">>
$ detector path/to/executables -d bintropy --binary
<<snipped ; in this case, as Bintropy only supports binary classification, --binary is necessary>>
Machine Learning models are fine-tuned through the YAML configuration files. They consist in:
algorithms.yml
: the algorithms that are used with their static or dynamic parameters while training modelsfeatures.yml
: the characteristics to be considered while training and using modelsThe PREPARE phase, especially feature engineering, is fine-tuned with the features YAML definition. Note that feature extraction is achieved with the pbox
package of the Packing-Box while feature derivation and transformation is fine-tuned via the features YAML file.
The TRAIN phase is fine-tuned through the algorithms YAML file by setting the static and/or cross-validation parameters.
The PREPARE phase, especially dataset generation, is achieved with the dataset
tool.
Operation | Description | Command |
---|---|---|
make ![]() |
Make a new dataset, either fully packed or mixed with not-packed samples | # dataset make dataset -c PE -n 200 -s /path/to/pe |
merge ![]() |
Merge two datasets | # dataset merge dataset dataset2 |
select ![]() |
Select a subset of a dataset to create a new one | # dataset select dataset dataset2 -q "format == 'PE32'" |
update ![]() |
Update a dataset with new samples given their labels | # dataset update dataset -l labels.json -s folder-of-executables |
The VISUALIZE phase can be performed with the dataset
and visualizer
tools.
In order to visualize feature values:
$ dataset plot test-mix byte_0_after_ep byte_1_after_ep --multiclass
In order to visualize samples (aims to compare the not-packed and some packed versions):
$ visualizer plot "PsExec.exe$" dataset -s -l not-packed -l MEW -l RLPack -l UPX
This will work for instance for a structure formatted as such:
folder/
+-- not-packed/PsExec.exe
+-- packed
+-- MEW/mew_PsExec.exe
+-- RLPack/rlpack_PsExec.exe
+-- UPX/upx_PsExec.exe
The TRAIN and PREDICT phases of the pipeline are achieved with the model
tool.
Operation | Description | Command |
---|---|---|
compare ![]() |
Compare the performance metrics of multiple models | # model compare model --dataset dataset --model model2 |
test ![]() |
Test a model on a given dataset | # model test model --name dataset |
train ![]() |
Train a model given an algorithm and input dataset | # model train dataset --algorithm dt |
You may also like these: