Use a RL model to select diversification mode

Environment Setting

Please install IDA Pro (7.3), bindiff, sqlite3, perf, tensorflow(>=2.0) and keras first.

To install the python2 environment for uroboros, please refer to the uroboros-diversification/README.md

Please modify values in config.json

uroboros_env is the python virtualenv for running uroboros $uroboros_env/bin/python must exist

prj_dir is the directory for this project

ida_pro_path is the executable of ida pro. idat64 is tested.

bindiff_path is the executable of bindiff, the default path of this is /opt/bindiff/bin/bindiff

perf_path is the executabtle of perf, the default path of this is /usr/bin/perf

If modify this json file does not work, you could modify src/utils/config.py directly.

Please modify the py in run_keras.sh.

Build a Coreutil Case to Train

./build_coreutils.sh [bin_name]

Please see provided cases in ./samples

Training Model

sh ./run_with_screen.sh keras

This will need root right, because we use perf to evaluate the running time of a program.

You need to make sure there is at least 25GB space on current disk to save checkpoints, or you can modify the checkpoints_dir parameter in keras_rl.py

Using Models

The models are saved on disk with file names dqn_model_[iteration]. Use your keras environment to run

python3 ./keras_load_and_obfs.py [model_path]

Please note the above command will clean contents in the output directory. See the created coreutils_callable_env.py. The default output directory is ./output

Explanation

the data is in ./explanation, and these data comes from running ./obfs_explainer.py and ./read_html_explanation.py

Test other binaries

To run cases not provided, please look into file coreutils_env_config.py

Please modify these directories variables in coreutils_env_config.py

src_coreutils_path and output_dir, the binary being tested must exist in src_coreutils_path, and train_bin_dict should add an item.

key is the name of this binary, and value is a tuple of

a.the command to be tested;

b.the time cost of this command. You could initial it with None, or write it manually. Use sudo perf stat -r [repeat-times] -e task-clock [your command] to get it.

python coreutils_env_config.py gym2 coreutils_callable_env, a file coreutils_callable_env.py will be created. You could change the file name, but change the import in keras_rl.py at the same time.