Python symbolic execution package
Project description
SEMA - ToolChain using Symbolic Execution for Malware Analysis.
██████ ▓█████ ███▄ ▄███▓ ▄▄▄
▒██ ▒ ▓█ ▀ ▓██▒▀█▀ ██▒▒████▄
░ ▓██▄ ▒███ ▓██ ▓██░▒██ ▀█▄
▒ ██▒▒▓█ ▄ ▒██ ▒██ ░██▄▄▄▄██
▒██████▒▒░▒████▒▒██▒ ░██▒ ▓█ ▓██▒
▒ ▒▓▒ ▒ ░░░ ▒░ ░░ ▒░ ░ ░ ▒▒ ▓▒█░
░ ░▒ ░ ░ ░ ░ ░░ ░ ░ ▒ ▒▒ ░
░ ░ ░ ░ ░ ░ ░ ▒
░ ░ ░ ░ ░ ░
Table of Contents
- Architecture
- Recommended Installation and Usage
- Dockerhub Installation and Usage
- Pypi Installation and Usage
- Credentials
Architecture
Toolchain architecture
Our toolchain is represented in the following figure and works as follows:
- A collection of labelled binaries from different malware families is collected and used as the input of the toolchain.
- Angr, a framework for symbolic execution, is used to execute binaries symbolically and extract execution traces. For this purpose, different heuristics have been developed to optimize symbolic execution.
- Several execution traces (i.e., API calls used and their arguments) corresponding to one binary are extracted with Angr and gathered together using several graph heuristics to construct a SCDG.
- These resulting SCDGs are then used as input to graph mining to extract common graphs between SCDGs of the same family and create a signature.
- Finally, when a new sample has to be classified, its SCDG is built and compared with SCDGs of known families using a simple similarity metric.
This repository contains a first version of a SCDG extractor. During the symbolic analysis of a binary, all system calls and their arguments found are recorded. After some stop conditions for symbolic analysis, a graph is built as follows: Nodes are system calls recorded, edges show that some arguments are shared between calls.
When a new sample has to be evaluated, its SCDG is first built as described previously. Then, gspan
is applied to extract the biggest common subgraph and a similarity score is evaluated to decide if the graph is considered as part of the family or not. The similarity score S
between graph G'
and G''
is computed as follows:
Since G''
is a subgraph of G'
, this is calculating how much G'
appears in G''
.
Another classifier we use is the Support Vector Machine (SVM
) with INRIA graph kernel or the Weisfeiler-Lehman extension graph kernel.
A web application is available and is called SemaWebApp. It allows to manage the launch of experiments on SemaSCDG and/or SemaClassifier.
Main depencies:
* Python 3.8
* Docker >=26.1.3 , docker buildx, Docker Compose >=v2.27.0
* radare2
* libvirt-dev, libgraphviz-dev, wheel
Interesting links
Extracting database
To extract the database, use the following commands:
cd databases/Binaries
./extract_deploy_db.sh
Password for archive is "infected". Warning : it contains real samples of malwares.
Compressing database
To compress the database, use the following commands:
#To zip back the test database
cd databases/Binaries
./compress_db.sh
Pypi installation and usage
To use the toolchain without docker container by using the Pypi package to install dependencies, use :
pip install sema-toolchain
After cloning the git you can then use the toolchain without docker
Pypy3 usage
Pypy3 can be used to launch experiments, make sure to install pypy3 :
sudo add-apt-repository ppa:pypy/ppa
sudo apt update
sudo apt install pypy3
Then install the dependecies on pypy3 :
pypy3 -m pip install -r /sema_scdg/requirements_pypy.txt
How to use ?
Use SemaSCDG
To run experiments, run :
python3 sema_scdg/application/SemaSCDG.py sema_scdg/application/configs/config.ini
Or if you want to use pypy3:
pypy3 sema_scdg/application/SemaSCDG.py sema_scdg/application/configs/config.ini
Configuration files
The parameters are put in a configuration file : configs/config.ini
. Feel free to modify it or create new configuration files to run different experiments.
The output of the SCDG are put into database/SCDG/runs/
by default. If you are not using volumes and want to save some runs from the container to your host machine, use :
make save-scdg-runs ARGS=PATH
Parameters description
SCDG module arguments
expl_method:
DFS Depth First Search
BFS Breadth First Search
CDFS Coverage Depth-First Search Strategy (Default)
CBFS Coverage Breadth First Search
graph_output:
gs .GS format
json .JSON format
EMPTY if left empty then build on all available format
packing_type:
symbion Concolic unpacking method (linux | windows [in progress])
unipacker Emulation unpacking method (windows only)
SCDG exploration techniques parameters:
jump_it Number of iteration allowed for a symbolic loop (default : 3)
max_in_pause_stach Number of states allowed in pause stash (default : 200)
max_step Maximum number of steps allowed for a state (default : 50 000)
max_end_state Number of deadended state required to stop (default : 600)
max_simul_state Number of simultaneous states we explore with simulation manager (default : 5)
Binary parameters:
n_args Number of symbolic arguments given to the binary (default : 0)
loop_counter_concrete How many times a loop can loop (default : 10240)
count_block_enable Enable the count of visited blocks and instructions
sim_file Create SimFile
entry_addr Entry address of the binary
SCDG creation parameter:
min_size Minimum size required for a trace to be used in SCDG (default : 3)
disjoint_union Do we merge traces or use disjoint union ? (default : merge)
not_comp_args Do we compare arguments to add new nodes when building graph ? (default : comparison enabled)
three_edges Do we use the three-edges strategy ? (default : False)
not_ignore_zero Do we ignore zero when building graph ? (default : Discard zero)
keep_inter_SCDG Keep intermediate SCDG in file (default : False)
eval_time TODO
Global parameter:
concrete_target_is_local Use a local GDB server instead of using cuckoo (default : False)
print_syscall Print the syscall found
csv_file Name of the csv to save the experiment data
plugin_enable Enable the plugins set to true in the config.ini file
approximate Symbolic approximation
is_packed Is the binary packed ? (default : False, not yet supported)
timeout Timeout in seconds before ending extraction (default : 600)
string_resolve Do we try to resolv references of string (default : True)
log_level Level of log, can be INFO, DEBUG, WARNING, ERROR (default : INFO)
family Family of the malware (default : Unknown)
exp_dir Name of the directory to save SCDG extracted (default : Default)
binary_path Relative path to the binary or directory (has to be in the database folder)
fast_main Jump directly into the main function
Plugins:
plugin_env_var Enable the env_var plugin
plugin_locale_info Enable the locale_info plugin
plugin_resources Enable the resources plugin
plugin_widechar Enable the widechar plugin
plugin_registery Enable the registery plugin
plugin_atom Enable the atom plugin
plugin_thread Enable the thread plugin
plugin_track_command Enable the track_command plugin
plugin_ioc_report Enable the ioc_report plugin
plugin_hooks Enable the hooks plugin
To know the details of the angr options see Angr documentation
You also have a script MergeGspan.py
in sema_scdg/application/helper
which could merge all .gs
from a directory into only one file.
Run multiple experiments automatically
If you wish to run multiple experiments with different configuration files, the script multiple_experiments.sh
is available. When being in the folder containing SemaSCDG.py :
# To show usage
./multiple_experiments.sh -h
# Run example
./multiple_experiments.sh -m python3 -c configs/config1.ini configs/config2.ini
Tests
To run the test :
python3 scdg_tests.py test_data/config_test.ini
Tutorial
There is a jupyter notebook providing a tutorial on how to use the scdg. To launch it, run
jupyter notebook --ip=0.0.0.0 --port=5001 --no-browser --allow-root --NotebookApp.token=''
and visit http://127.0.0.1:5001/tree
on your browser. Go to /Tutorial
and open the jupyter notebook.
Use SemaClassifier
Just run the script :
python3 SemaClassifier.py FOLDER/FILE
usage: update_readme_usage.py [-h] [--threshold THRESHOLD] [--biggest_subgraph BIGGEST_SUBGRAPH] [--support SUPPORT] [--ctimeout CTIMEOUT] [--epoch EPOCH] [--sepoch SEPOCH]
[--data_scale DATA_SCALE] [--vector_size VECTOR_SIZE] [--batch_size BATCH_SIZE] (--classification | --detection) (--wl | --inria | --dl | --gspan)
[--bancteian] [--delf] [--FeakerStealer] [--gandcrab] [--ircbot] [--lamer] [--nitol] [--RedLineStealer] [--sfone] [--sillyp2p] [--simbot]
[--Sodinokibi] [--sytro] [--upatre] [--wabot] [--RemcosRAT] [--verbose_classifier] [--train] [--nthread NTHREAD]
binaries
Classification module arguments
optional arguments:
-h, --help show this help message and exit
--classification By malware family
--detection Cleanware vs Malware
--wl TODO
--inria TODO
--dl TODO
--gspan TODOe
Global classifiers parameters:
--threshold THRESHOLD
Threshold used for the classifier [0..1] (default : 0.45)
Gspan options:
--biggest_subgraph BIGGEST_SUBGRAPH
Biggest subgraph consider for Gspan (default: 5)
--support SUPPORT Support used for the gpsan classifier [0..1] (default : 0.75)
--ctimeout CTIMEOUT Timeout for gspan classifier (default : 3sec)
Deep Learning options:
--epoch EPOCH Only for deep learning model: number of epoch (default: 5) Always 1 for FL model
--sepoch SEPOCH Only for deep learning model: starting epoch (default: 1)
--data_scale DATA_SCALE
Only for deep learning model: data scale value (default: 0.9)
--vector_size VECTOR_SIZE
Only for deep learning model: Size of the vector used (default: 4)
--batch_size BATCH_SIZE
Only for deep learning model: Batch size for the model (default: 1)
Malware familly:
--bancteian
--delf
--FeakerStealer
--gandcrab
--ircbot
--lamer
--nitol
--RedLineStealer
--sfone
--sillyp2p
--simbot
--Sodinokibi
--sytro
--upatre
--wabot
--RemcosRAT
Global parameter:
--verbose_classifier Verbose output during train/classification (default : False)
--train Launch training process, else classify/detect new sample with previously computed model
--nthread NTHREAD Number of thread used (default: max)
binaries Name of the folder containing binary'signatures to analyze (Default: output/save-SCDG/, only that for ToolChain)
Example
This will train models for input dataset
python3 SemaClassifier.py --train output/save-SCDG/
This will classify input dataset based on previously computed models
python3 SemaClassifier.py output/test-set/
Tests
To run the classifier tests :
python3 classifier_tests.py configs/config_test.ini
Credentials
Main authors of the projects:
-
Charles-Henry Bertrand Van Ouytsel (UCLouvain)
-
Christophe Crochet (UCLouvain)
-
Khanh Huu The Dam (UCLouvain)
-
Oreins Manon (UCLouvain)
Under the supervision and with the support of Fabrizio Biondi (Avast)
Under the supervision and with the support of our professor Axel Legay (UCLouvain) (:heart:)
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
File details
Details for the file sema_toolchain-0.0.11.tar.gz
.
File metadata
- Download URL: sema_toolchain-0.0.11.tar.gz
- Upload date:
- Size: 9.5 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/5.0.0 CPython/3.10.12
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 838ca0287e497ce7b938f1a2fafdb2bc9933141eccab3584b639aa6d6c5fd576 |
|
MD5 | 8af44bb80bbe4074219bd4be34f52118 |
|
BLAKE2b-256 | 8530031de95490470d18e522643000b04498edf89c9ffec556f534c3be599e9b |
File details
Details for the file sema_toolchain-0.0.11-py3-none-any.whl
.
File metadata
- Download URL: sema_toolchain-0.0.11-py3-none-any.whl
- Upload date:
- Size: 8.2 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/5.0.0 CPython/3.10.12
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | f1e565afc33dec0f2c821c45d934d60a0e9d7a54814e5aa0c673812a10d2e7e9 |
|
MD5 | c2de36b2d5442d883c70fe16e447ac0a |
|
BLAKE2b-256 | 9a30bbedfe0677825e6980f01d2110ae3e7db50a8980eb2f1c7f5d214126fc86 |