It makes unicorn traces.
Using Unicorn as a basis, Rainbow aims to provide an easy scripting interface to loosely emulate embedded binaries, trace them to perform side-channels, and simulate fault injections.
This is to allow quick and easy testing of physical attack resistance of code snippets, in order to help developers have a first evaluation of the resistance of their code.
An introduction is available here.
A blogpost demonstrating how to turn this tool into an automatic fault injection test pipeline is here, with the corresponding Rust code here.
You will need Python 3.7 at least.
pip install .
If Unicorn or Capstone fails to install somehow:
- Unicorn: http://www.unicorn-engine.org/download/
- Capstone: https://www.capstone-engine.org/
For the side-channel examples, you need to the latest Lascar, the following command installs the necessary packages.
pip install .[examples]
If you wish to use another version of Python, you can drop an issue and we will look into it.
Some examples will use Lascar's side-channel attacks and try to display traces using a custom plotter (visplot) built on top of Vispy. If you want to run those, you will need Vispy and pyqt5
for the instruction trace + execution trace viewer.
In the ./examples/
folder, you will find:
- x64_pimpmyxor.py: basic emulation of this challenge
- CortexM_AES: a simple ARM Thumb AES
- Hacklu2009: a side-channel solution of a whitebox challenge
- HW_analysis: a side-channel simulation of a pin comparison, and a fault injection simulation
- ledger_ctf2: side-channel solution of a whitebox challenge
- OAES: an x86 whitebox tracing example that discards useless instructions
- SecAESSTM32: a starting point to test ANSSI's STM32 secure AES implementation
See the x64_pimp_my_xor
example for a debug trace.
In the comment part of each line (after the semicolon), the memory access that was performed is written in a simplified way: [address] <- value
for a load or value -> [address]
for a store. Right after, if any register was modified during this instruction, its new value is shown.
At a branch instruction, if the destination is a known function, its name is shown together with the return address and the function's address.
Grab a device or generic emulator like so:
from rainbow.devices import rainbow_stm32f215
e = rainbow_stm32f215()
Load a binary:
e.load('file', typ='.elf')
e.setup()
File type is guessed on the extension when possible (.elf, .hex).
Starting the emulation is done like so:
e.start(start_address, stop_address, count=number_of_instructions)
Just like with unicorn. The underlying Unicorn instance is always available as e.emu
.
To enable printing as code gets executed, simply use the Print
flag.
from rainbow import Print
import colorama
colorama.init() # Only do this once to enable colors
e = rainbow_stm32f215(print=Print.Code | Print.Functions) # see other values of the flag
Rainbow only produces an execution trace, without applying any processing (such as adding noise) on the values. This is left as some post-processing, so that the user can apply its own leakage model and simulate various conditions from the same traces. Also, not introducing any noise allows testing in a worst-case scenario, which can yield important results.
To perform the analysis, one can use Lascar.
You can find some scripts in the examples
folder here which already use it.
To setup tracing (to produce an execution trace) use the trace_config
option
to the emulator. The following piece of code sets up tracing of register
using the Hamming weight leakage model.
from rainbow import TraceConfig, HammingWeight
e = rainbow_stm32f215(trace_config=TraceConfig(register=HammingWeight()))
e.load('file', typ='.elf')
e.setup()
e.start(start_address)
print(e.trace)
# [{"type": "code", "register": 7}, {"type": "code": "register": 5}]
If you setup tracing for mem_address
, then the e.trace
list will have dictionaries
like {"type": "mem_read", "address": 1234}
or {"type": "mem_write", "address": 1234}
with the value of the address
entry passed through the leakage model. Tracing for
mem_value
does the same, but traces memory values read or written and produces entries
like {"type": "mem_read", "value": 1234}
. Note that these approaches can be combined,
resulting in the dictionary having both an address
and value
entries.
If you setup tracing for code
, dissasembled instructions will be available in the
trace with dictionaries like {"type": "code", "instruction": " 404 ldm.w r0, {r4, r5, r6, r7}"}
.
Note that this tracing option combined with register tracing produces a dictionary with
both instruction
and register
entries.
In the case of hardware wallets for example, one could check that:
- The PIN verification procedure does not allow to use a bad password even with a controlled instruction skip
- The scalar multiplication procedure does not leak any information on the used scalar
- a purely software AES is protected against basic DPA attacks without using lab testing equipment (oscilloscope, current/EM probes, ...)
Rainbow and Lascar allow testing implemented countermeasures were correctly coded and the compiler did not interfere. It cannot, however, verify against hardware-related leaks such as some sequence of operations that somehow cancels out random masks on a bus or hidden register.
Whiteboxed encryption primitives could also be broken using this tool, instead of e.g. Intel Pin or Valgrind to trace execution. Unicorn has several advantages in this regard:
- Can be used on a different platform than that of the target binary
- Allows easy manipulation of the state (for example redefining an external call to
rand()
in Python)
Disadvantages:
- Some reverse engineering necessary !
As a whitebox example (available in examples/OAES
, below is the result of the variance of SECCON 2016's OAES encryption function, which has a heavy control flow obfuscation.
One can clearly see the 10 rounds of the AES despite this obfuscation:
Embedded devices:
- STM32F215
- STM32l431
Generic emulators:
- ARM
- ARM Cortex M
- x86
- x86_64
- M68K
File formats:
- ELF
- Intel Hex file
- PE