QBinDiff
QBinDiff is an experimental binary diffing tool addressing the diffing as a Network Alignement Quadratic Problem.
But why developing yet another differ when Bindiff works well?
Bindiff is great, no doubt about it, but we have no control on the diffing process. Also, it works great on standard binaries but it lacks flexibility on some corner-cases (embedded firmwares, diffing two portions of the same binary etc...).
A key idea of QBinDiff is enabling tuning the diffing programmatically by:
- writing its own feature
- being able to enforce some matches
- emphasizing either on the content of functions (similarity) or the links between them (callgraph)
In essence, the idea is to be able to diff by defining its own criteria which sometimes, is not the control-flow and instructions but could for instance, be data-oriented.
Last, QBinDiff as primarily been designed with the binary-diffing use-case in mind, but it can be applied to various other use-cases like social-networks. Indeed, diffing two programs boils down to determining the best alignment of the call graph following some similarity criterion.
Indeed, solving this problem is APX-hard, that why QBinDiff uses a machine learning approach (more precisely optimization) to approximate the best match.
Like Bindiff, QBinDiff also works using an exported disassembly of program obtained from IDA. Originally using BinExport, it now also support Quokka as backend, which extracted files, are more exhaustive and also more compact on disk (good for large binary dataset).
Note
QBinDiff is an experimental tool for power-user where many parameters, features, thresholds or weights can be adjusted. Obtaining good results usually requires tuning these parameters.
(Please note that QBinDiff does not intend to be faster than other differs, but rather being more flexible.)
Warning
QBinDiff graph alignment is very memory intensive (compute large matrices), it can fill RAM if not cautious. Try not diffing binaries larger than +10k functions. For large program use very high sparsity ratio (0.99).
Documentation
The documentation can be found on the diffing portal or can be manually built with
pip install .[doc]
cd doc
make html
Below you will find some sections extracted from the documentation. Please refer to the full documentation in case of issues.
Installation
QBinDiff can be installed through pip with:
pip install qbindiff
As some part of the algorithm are very CPU intensive the installation will compile some components written in native C/C++.
As depicted above, QBinDiff relies on some projects (also developed at Quarkslab):
- python-binexport, wrapper on the BinExport protobuf format.
- python-bindiff, wrapper around bindiff (used to write results as Bindiff databases)
- Quokka, another binary exported based on IDA. Faster than binexport and more exhaustive (thus diffing more relevant)
Usage (command line)
After installation, the binary qbindiff is available in the path.
It takes in input two exported files and start the diffing analysis. The result can then
be exported in a BinDiff file format.
The default format for input files is BinExport,
for a complete list of backend loader look at the -l1, --loader1 option in the help.
The complete command line options are:
Usage: qbindiff [OPTIONS] <primary file> <secondary file>
QBinDiff is an experimental binary diffing tool based on machine learning technics, namely Belief propagation.
Options:
-l1, --loader1 <loader> Loader type to be used for the primary. Must be one of these ['binexport', 'quokka',
'ida'] [default: binexport]
-l2, --loader2 <loader> Loader type to be used for the secondary. Must be one of these ['binexport', 'quokka',
'ida'] [default: binexport]
-f, --feature <feature> Features to use for the binary analysis, it can be specified multiple times.
Features may be weighted by a positive value (default 1.0) and/or compared with a
specific distance (by default the option -d is used) like this <feature>:<weight>:<distance>.
For a list of all the features available see --list-features.
-n, --normalize Normalize the Call Graph (can potentially lead to a partial matching). [default
disabled]
-d, --distance <function> The following distances are available ['canberra', 'euclidean', 'cosine',
'jaccard_strong'] [default: canberra]
-s, --sparsity-ratio FLOAT Ratio of least probable matches to ignore. Between 0.0 (nothing is ignored) to 1.0
(only perfect matches are considered) [default: 0.75]
-sr, --sparse-row Whether to build the sparse similarity matrix considering its entirety or processing
it row per row
-t, --tradeoff FLOAT Tradeoff between function content (near 1.0) and call-graph information (near 0.0)
[default: 0.75]
-e, --epsilon FLOAT Relaxation parameter to enforce convergence [default: 0.5]
-i, --maxiter INTEGER Maximum number of iteration for belief propagation [default: 1000]
-e1, --executable1 PATH Path to the primary raw executable. Must be provided if using quokka loader
-e2, --executable2 PATH Path to the secondary raw executable. Must be provided if using quokka loader
-o, --output PATH Write output to PATH
-ff, --file-format [bindiff] The file format of the output file. Supported formats are [bindiff] [default:
bindiff]
-v, --verbose Activate debugging messages. Can be supplied multiple times to increase verbosity
--version Show the version and exit.
--arch-primary TEXT Force the architecture when disassembling for the primary. Format is
'CS_ARCH_X:CS_MODE_Ya,CS_MODE_Yb,...'
--arch-secondary TEXT Force the architecture when disassembling for the secondary. Format is
'CS_ARCH_X:CS_MODE_Ya,CS_MODE_Yb,...'
--list-features List all the available features
-h, --help Show this message and exit.
Library usage
The strength of qBinDiff is to be usable as a python library. The following snippet shows an example of loading to binexport files and to compare them using the mnemonic feature.
from qbindiff import QBinDiff, Program
from qbindiff.features import WeisfeilerLehman
from pathlib import Path
p1 = Program(Path("primary.BinExport"))
p2 = Program(Path("secondary.BinExport"))
differ = QBinDiff(p1, p2)
differ.register_feature_extractor(WeisfeilerLehman, 1.0, distance='cosine')
differ.process()
mapping = differ.compute_matching()
output = {(match.primary.addr, match.secondary.addr) for match in mapping}Contributing & Contributors
Any help, or feedback is greatly appreciated via Github issues, pull requests.
Current:
- Robin David
- Riccardo Mori
- Roxane Cohen
Past:
- Alexis Challande
- Elie Mengin