cpu_rec

Description

cpu_rec is a tool that recognizes cpu instructions in an arbitrary binary file. It can be used as a standalone tool, or as a plugin for binwalk (https://github.com/devttys0/binwalk).

Installation instructions

Standalone tool

1. Copy cpu_rec.py and cpu_rec_corpus in the same directory.
2. If you don't have the lzma module installed for your python (this tool works either with python3 or with python2 >= 2.4) then you should unxz the corpus files in cpu_rec_corpus.
3. If you want to enhance the corpus, you can add new data in the corpus directory. If you want to create your own corpus, please look at the method build_default_corpus in the source code.

For use as a binwalk module

Same as above, but the installation directory must be the binwalk module directory: $HOME/.config/binwalk/modules.

You'll need a recent version of binwalk, that includes the patch provided by https://github.com/devttys0/binwalk/pull/241 .

How to use the tool

As a binwalk module

Add the flag -% when using binwalk.

Be patient. Waiting a few minutes for the result is to be expected. On my laptop the tool takes 25 seconds and 1 Gb of RAM to create the signatures for 70 architectures, and then the analysis of a binary takes one minute per Mb. If you want the tool to be faster, you can remove some architectures, if you know that your binary is not one of them (typically Cray or MMIX are not found in a firmware).

As a standalone tool

Just run the tool, with the binary file(s) to analyze as argument(s) The tool will try to match an architecture for the whole file, and then to detect the largest binary chunk that corresponds to a CPU architecture; usually it is the right answer, but one should not forget that this tool is heuristic and that some binary files contain instructions for multiple architectures, therefore a more detailed analysis may be needed.

If the result is not satisfying, prepending twice -v to the arguments makes the tool very verbose; this is helpful when adding a new architecture to the corpus or when there are doubts on the raw result of the tool.

If https://github.com/LRGH/elfesteem is installed, then the tool also extract the text section from ELF, PE, Mach-O or COFF files, and outputs the architecture corresponding to this section; the possibility of extracting the text section is also used when building a corpus from full binary files.

If https://lief-project.github.io/ is installed, then the tool can use it instead of elfesteem. If both are installed, option '-l' gives priority to lief.

Option -d followed by a directory dumps the corpus in that directory; using this option one can reconstruct the default corpus.

As a python module

The function which_arch takes a bytestring as input and outputs the name of the architecture, or None. Loading the training data is done during the first call of which_arch, and calling which_arch with no argument does this precomputation only.

For example

>>> from cpu_rec import which_arch
>>> which_arch()
>>> which_arch(b'toto')
>>> which_arch(open('/bin/sh').read())
'X86-64'

Create a corpus or extend the existing corpus

Each architecture is defined by a file in cpu_rec_corpus. Only file names ending with .corpus, which can be compressed with xz.

The corpus file shall contain instructions for the target architecture. As you can see in build_default_corpus, most of the default corpus has been created by extracting the TEXT section of an executable.

If you want to add an new architecture (e.g. 78k as described below) then you have to find a binary, and extract the executable section (the command line to extract the 78k code from the Metz firmware is dd if=MB50AF1_NikonV12.bin of=Nec78k.corpus bs=1 skip=0x2ba count=0x7d5a).

Examples

Running the tool as a binwalk module typically results in:

…

Important: it is usually a good idea to start the analysis of an unknown binary with some entropy analysis. cpu_rec assumes that it has been done, but to protect the user against overlooking this aspect, it displays the entropy. If the entropy value is above 0.9, it is probably encrypted or compressed data, and therefore the result of cpu_rec should be meaningless.

We can notice that during the analysis of ALPHA/NTDLL.DLL small chunks are wrongly detected as non-Alpha architectures. They should be ignored. But some files can contain multiple architectures, e.g. Mach-O/OSXII which is a Mach-O FAT file with ppc and i386 executables.

More documentation

The tool has been presented at SSTIC 2017, with a full paper describing why this technique has been used for the recognition of architectures. A video of the presentation and the slides are available.

This presentation was made in French. A translation in English of the slides is available, a translation in English of the paper is in progress.

Known architectures in the default corpus

6502 68HC08 68HC11 8051 Alpha ARC32eb ARC32el ARcompact ARM64 ARMeb ARMel ARMhf AVR AxisCris Blackfin Cell-SPU CLIPPER CompactRISC Cray CUDA Epiphany FR-V FR30 FT32 H8-300 H8S HP-Focus HP-PA i860 IA-64 IQ2000 M32C M32R M68k M88k MCore Mico32 MicroBlaze MIPS16 MIPSeb MIPSel MMIX MN10300 Moxie MSP430 NDS32 NIOS-II OCaml PDP-11 PIC10 PIC16 PIC18 PIC24 PPCeb PPCel RISC-V RL78 ROMP RX S-390 SPARC STM8 Stormy16 SuperH TILEPro TLCS-90 TMS320C2x TMS320C6x TriMedia V850 VAX Visium WASM WE32000 X86-64 X86 Xtensa Z80 #6502#cc65

Because of licencing issues, the following architectures are not in the default corpus, but they can be manually added: 78k TriCore

Licence

The tool

The cpu_rec.py file is licenced under a Apache Licence, Version 2.0.

The default corpus

The files in the default corpus have been built from various sources. The corpus is a collection of various compressed files, each compressed file is dedicated to the recognition of one architecture and is made by the compression of the concatenation of one or many binary chunks, which come from various origins and have various licences. Therefore, the default corpus is a composite document, each sub-document (the chunk) being redistributed under the appropriate licence.

The origin of each chunk is described in cpu_rec.py, in the function build_default_corpus. The licences are:

• files libgmp.so, libc.so, libm.so come from Debian binary distributions and are distributed under GPLv2 (and LGPLv3 for recent versions of libgmp) and the source code is available from http://archive.debian.org/.
• busybox binaries come from https://busybox.net/downloads/binaries/ and are distributed under GPLv2.
• C-Kermit binaries come from ftp://kermit.columbia.edu/kermit/bin/ and are distributed under GPLv2 (according to ftp://kermit.columbia.edu/kermit/archives/COPYING but the status of each binary is not always clear).
• all files identified in build_default_corpus as part of the CROSS_COMPILED subdirectory have been built by myself. The corresponding source code are zlib (from http://zlib.net/, distributed under the zlib licence) or libjpeg (from http://www.ijg.org/, distributed under an unknown licence) or some other code based on public sources (e.g. https://anonscm.debian.org/cgit/pkg-games/bsdgames.git/tree/arithmetic/arithmetic.c modified to work with SDCC compilers).
• The camlp4 binary is built from https://github.com/ocaml/camlp4 and distributed under LGPLv2.
• The binary for TMS320C2x comes from https://github.com/slavaprokopiy/Mini-TMS320C28346/blob/master/For_user/C28346_Load_Program_to_Flash/Debug/C28346_Load_Program_to_Flash.out where it is distributed under an unknown licence.
• The binary for RISC-V comes from https://riscv.org/software-tools/ distributed under GPLv2 and can downloaded at https://github.com/radare/radare2-regressions/blob/master/bins/elf/analysis/guess-number-riscv64
• The binaries for PIC10 and PIC16 come from http://www.pic24.ru/doku.php/en/osa/ref/examples/intro where they are distributed under an unknown licence.
• The binary for PIC18 comes from https://github.com/radare/radare2-regressions/blob/master/bins/pic18c/FreeRTOS-pic18c.hex where it seems to be distributed under GPLv3 (or later).
• The binary for PIC24 comes from https://raw.githubusercontent.com/mikebdp2/Bus_Pirate/master/package_latest/BPv4/firmware/bpv4_fw7.0_opt0_18092016.hex distributed under Creative Commons Zero.
• The binary for 6502 comes from https://raw.githubuserco