- Docker build support!
- Add bug list
- Goshawk now supports Clang-15.0.0
To see bugs found by Goshawk, visit bug_list page. You can also see some details about the bugs found by Goshawk.
data_process:The scripts for pre-processing, parsing and normalizing the function prototypes.model:Pre-trained Siamese network, can be directly used to classify functions.plugins:Clang and CSA plugins used by Goshawk.plugins_src:The source codes of Clang plugins.subword_dataset:The learned subword vocabulary and embedding for function prototype segmentation, and official MM function list.
run.py:The entry point of Goshawk, performs each steps of Goshawk.train.py:Train the Siamese network.cal_metric.py:Evaluate the accuracy of the trained model.similarity_inference.py:Utilize the trained Siamese network to generate similarity scores for each function prototype.mysegment.py:The ULM based function prototype segmentation algorithm.frontend_checker.py:Validate the MM functions according the function prototype and data flow.
Directly use our image released on DockerHub:
docker pull mmmiracle/goshawk
Or build docker image by yourself:
docker build -t goshawk .
robin-map
python 3.7+
tensorflow = 2.1
CodeChecker
Clang v15.0.0
Download the subword embeddings to the directory subword_dataset/word_embedding.
You can install robin-map from https://github.com/Tessil/robin-map.
You can install CodeChecker from https://github.com/Ericsson/codechecker.
You can download the version of Clang-15.0.0 form this link, or compile a clang-15.0.0 by yourself.
Before using this tool, you need to record the compilation commands used by each file to compile the source code of the project, and then the further analysis will be based on these compilation commands.
We can use CodeChecker to record the required compilation commands. For the projects which use Makefile to build,
we can use the log -b cmd to encapsulate the make related cmd to record the compiling process:
export CC=clang
export CXX=clang++
CodeChecker log -b "make CC=clang HOSTCC=clang -j$(nproc)" -o compilation.json
Remember that set your default compilers to clang and clang++.
The compilation commands will be recorded in the file of compilation.json.
note: For large project, like linux kernel, you should guarantee that there is at least 300GB ROM on you hard disk.
Currently, you only need one command to analyze a project by Goshawk:
python3 run.py target_project_path
But you should make sure that there is a compilation.json file of your project under the target_project_path.
The MM functions and their corresponding MOSs will be generated at output/alloc and output/free.
Note: All the MM functions in
AllocNormalFile.txt,AllocCustomFile.txt,FreeNormalFile.txtandFreeCustomFile.txtare considered as customized. They are separated as *NormalFile.txt and *CustomFile.txt because all MM functions in *NormalFile.txt behavior likemallocorfree. It is an implementation decison.
The bug detection results will be generated at output/report_html/index.html.
Function normalize_prototype_file(in_file, out_file) in normalize.py can be used to segment function prototypes.
It Segments and normalizes the function prototypes in the in_file, and the results are saved at out_file.
For example,
before: void * kmalloc_array(size_t n, size_t size, gfp_t flags)
after: <cls> <ptr> kmalloc array ( <noptr> n <dot> <noptr> size <dot> <noptr> flags )
Ⅲ.B Re-train Simaese network for your customized target function identification task (e.g.,MM functions, crypto functions,...).
Take crypto function as example, the dataset should be the prototypes of your collected crypto functions. Each line is a function prototype.
crypto.txt
-------------
int crypto_aead_encrypt(struct aead_request *req)
int crypto_aead_decrypt(struct aead_request *req)
static int crypto_aegis128_encrypt_generic(struct aead_request *req)
static int crypto_aegis128_decrypt_simd(struct aead_request *req)
void crypto_aegis128_encrypt_chunk_neon(void *state, void *dst, const void *src,unsigned int size)
void crypto_aegis128_decrypt_chunk_neon(void *state, void *dst, const void *src,unsigned int size)
static int crypto_authenc_esn_encrypt(struct aead_request *req)
static int crypto_authenc_esn_decrypt(struct aead_request *req)
...
We have implemented the re-train of Siamese network in the script Re-train.py.
It takes two arguments:
- training corpus
- your trained model name
For example:
python Re-train.py crypto.txt crypto
After the training finished, your trained model which names "crypto" is saved at directory "model/crypto".
The already trained model were saved in the directory "model", you can use them to infer similarity for other functions directly.
We have implemented these functions in the script similarity_inference.py.
You can call the function similarity_inference(model_name, filename) to infer similarity
for the functions whose prototypes saved in the argument filename.
Here, model_name should be the name of model that save in the directory model.
For example, there is a file names test.func which contains the follow functions:
test.func
---------
void * mem_malloc(unsigned long size)
void mem_free(void *ptr)
void CAST_set_key(CAST_KEY *key, int len, const unsigned char *data)
We can call the function similarity_inference to infer similarities for them.
from similarity_inference import working_on_raw_function_prototype
similarity_inference("alloc", "test.func") # Infer the similarity for allocation functions.
The result are saved at "temp/func_name_similarity"
temp/func_name_similarity
----
mem_malloc 0.938829920833657
mem_free -0.9019584597976495
cast_set_key -0.9085114460471964
We release Goshawk source code in the hope of benefiting others. If you find this project useful, please consider citing:
@INPROCEEDINGS {Goshawk,
author = {Y. Lyu and Y. Fang and Y. Zhang and Q. Sun and S. Ma and E. Bertino and K. Lu and J. Li},
booktitle = {2022 2022 IEEE Symposium on Security and Privacy (SP) (SP)},
title = {Goshawk: Hunting Memory Corruptions via Structure-Aware and Object-Centric Memory Operation Synopsis},
year = {2022},
issn = {2375-1207},
pages = {1566-1566},
doi = {10.1109/SP46214.2022.00137},
url = {https://doi.ieeecomputersociety.org/10.1109/SP46214.2022.00137},
publisher = {IEEE Computer Society},
address = {Los Alamitos, CA, USA},
month = {may}
}
If your research work is inspired by or benefits from the NLP based function similarity inference module in Goshawk, please also consider citing:
@INPROCEEDINGS{SparrowHawk,
author = {Lyu, Yunlong and Gao, Wang and Ma, Siqi and Sun, Qibin and Li, Juanru},
title = {SparrowHawk: Memory Safety Flaw Detection via Data-Driven Source Code Annotation},
year = {2021},
isbn = {978-3-030-88322-5},
publisher = {Springer-Verlag},
address = {Berlin, Heidelberg},
url = {https://doi.org/10.1007/978-3-030-88323-2_7},
doi = {10.1007/978-3-030-88323-2_7},
booktitle = {Information Security and Cryptology: 17th International Conference, Inscrypt 2021, Virtual Event, August 12–14, 2021, Revised Selected Papers},
pages = {129–148},
numpages = {20},
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent