Neural Nets Can Learn Function Type Signatures From Binaries

Authors

EKLAVYA is designed by Zheng Leong Chua, Shiqi Shen, Prateek Saxena, Zhenkai Liang.

Dataset

The dataset available from this page is the collection of function type signatures, which includes function banaries, number of arguments and types. It is a good dataset for people who want to try learning techniques or heuristic approaches in binary analysis while spending less effort on collecting and preprocessing.

The dataset consists of three parts, which is described below:

binary.tar.gz. The compressed file saves 5168 binaries. These binaries are from 8 distinct packages: binutils, coreutils, findutils, sg3utils, utillinux, inetutils, diffutils, and usbutils. These 5168 binaries are obtained by using two commonly used compilers: gcc and clang, with different optimization levels ranging from O0 to O3 for both x86 and x64.
pickles.tar.gz. The compressed file saves the assembly code of each function and the ground truth of the function arguments by parsing the DWARF debug information.
clean_pickles.tar.gz. The compressed file saves the assembly code and the ground truth of the function arguments for sanitized functions. For this dataset, we removed the functions which are duplicates of other functions in the dataset. Given that the same piece of code compiled with different binaries will result in different offsets generated,we chose to remove all direct address used by instructions found in the function. For example, the instruction 'je 0x98' are represented as 'je '. After the substituion, we hash the function and remove functions with the same hashes. Other than duplicates, we also removed functions with less than four instructions as these small functions typically do not have any operation on arguments.

Function Representation

A function is represented as a dictionary having the following fields:

num_args: Integer - Number of arguments.
args_type: List - Type of each argument, as String, in the order they appear in the function declaration.
ret_type: String - Type of the value returned by the function.
inst_strings: List - The assembly code of each instruction composing the function, as strings.
inst_bytes: List - The bytecode of each instruction composing the function, as an array of values. (Each instruction is represented by one array of bytes.)
boundaries: Tuple (Integer, Integer) - The starting and ending address of the function.

Example:

FuncDict = {
    "num_args": 3,
    "args_type": ["int", "char", "struct structure1*"], 
    "ret_type": "int", 
    "inst_strings": ["mov eax, 1", "nop", "push 3"],
    "inst_bytes": [[0x01, 0x02], [0xff], [0x20, 0x30, 0x40]],
    "boundaries": (0x80010, 0x800f0)
}

Binary Representation

A binary saved in pickles.tar.gz and clean_pickles.tar.gz is represented as a Dict object, having the following fields:

functions: Dict - This dictionary contains as keys function names and as values function dictionaries which is described above. Dupilicate functions are removed from this dictionary in clean_pickles.tar.gz.
structures: Dict -This dictionary describing structures used in the binary files. Dictionary's keys are names of the structures and the values are lists containing the type of each field in the structure, in the order they are declared.
text_addr: String - Address of the text section in the binary.
binRawBytes: String - Raw contents of the binary file.
arch: String - Binary architecture.
binary_filename: String - Name of the binary elf file used to generate the binInfo.
function_calls: Dict - This dictionary containing information about function calls. A caller is described using its name and an array of instruction indices, starting from 0. (If 10 is in the array, the 10th instruction in the caller is a call to the callee.)
- Key: callee function's name
- Value: array of objects describing function callers

Example:

BinaryFileDict = {
    "functions": {
        "function1": funcDict1,
        "function2": funcDict2
    },
    "structures": {
        "structure1": ["int", "char [16]", "float"],
        "structure2": ["long", "long", "short"]
    },
    "text_addr": 0x800000,
    "binRawBytes": "\x00\x12\x22...",
    "arch": "i386",
    "binary_filename": "gcc-32-O1-coreutils-ls",
    "function_calls": {
         "func1": [
             {
                 # caller's name
                 "caller": "func2",
                 # the indices of calling instructions
                 "call_instr_indices": [10, 17, 29] 
             },
             {
                 "caller": "func2",
                 "call_instr_indices": [19]
             }
        ]
    }
}

Code

Requirements

tensorflow
numpy

Train Embedding Model

Prepare the input file for training embedding model

python prep_embed_input.py [options] -i input_folder

Link to prep_embed_input.py

input_folder: String - The data folder saving all binary information.

Options:

-o: String - The output file for training embedding model
-e: String - The file saving all error information
-m: String - The file saving the map information (int --> instruction (int list))

Train the embedding model

python train_embed.py -i input_path

Link to train_embed.py

input_path: String - The input file for training embedding model

Options:

-o: String - The output folder saving the trained embedding information
-tn: Integer - Number of threads
-sw: Integer - Saving frequency (Number of epochs). The trained information will be saved every several epochs.
-e: Integer - Dimension of the embedding vector for each instruction.
-ne: Integer - Number of epochs for training the embedding model
-l: Float - Learning rate
-nn: Integer - Number of negative samples
-b: Integer - Batch size
-ws: Integer - Window size
-mc: Integer - Ignoring all words with total frequency lower than this.
-s: Float - Subsampling threshold

Save the embedding vector

python save_embeddings.py [options] -p embed_pickle_path -m model_path

Link to save_embeddings.py

embed_pickle_path: String - The file saving all training parameters
model_path: String - The file saving the trained embedding model

Options:

-o: String: The output file saving the embedding vector for each instruction

Train RNN Model

python train.py [options] -d data_folder -o output_dir -f split_func_path -e embed_path

Link to train.py

data_folder: The folder saves the binary information.
output_dir: The directory is used to save the trained model & log information.
split_func_path: The file saves the training & testing function names.
embed_path: The file is the output file of save_embeddings.py, which saves the embedding vector of each instruction.

Options:

-t: Type of output labels. Possible value: num_args, type#0, type#1, ... (Default value: num_args)
- num_args: The trained model is used to predict the number of arguments for each function.
- type#0: The trained model is used to predict the type of first argument for each function.
- type#1: The trained model is used to predict the type of second argument for each function.
- ...
-dt: Type of input data. Possible value: caller and callee. (Default value: callee)
- caller: The input data is from caller.
- callee: The input data is function body.
-pn: Number of Processes. (Default value: 40)
-ed: Dimension of embedding vector for each instruction. (Default value: 256)
-ml: Maximum length of input sequences. (Default value: 500)
-nc: Number of Classes. (Default value: 16)
-en: Number of Epochs. (Default value: 100)
-s: The frequency for saving the trained model. If the value is 100, the trained model is going to be saved every 100 batches. (Default value: 100)
-do: Dropout value. (Default value: 0.8)
-nl: Number of layers in RNN. (Default value: 3)
-ms: Maximum number of model saved in the directory. (Default value: 100)
-b: Batch size. (Default value: 256)
-p: The frequency for showing the accuracy & cost value. (Default value: 20)

split_func_path Format

The split_func_path file saves the function names for training & testing dataset. If you are going to predict the type signatures from callees (function bodies), the function name is represented as "binary_file_name#func_name". If you are going to predict the type signatures from callers, the function name is represented as "binary_file_name#callee_name#caller_name#call_insn_indice".

Examples of split_func_path file for callees:

splitFuncDict = {
    'train':[
                'gcc-32-O1-binutils-objdump.pkl#OP_Rounding',
                'clang-32-O1-coreutils-csplit.pkl#keep_new_line',
                'gcc-32-O3-coreutils-mv.pkl#copy_internal',
                ...
            ],
    'test': [
                'gcc-32-O3-findutils-find.pkl#parse_amin',
                'gcc-32-O2-findutils-find.pkl#pred_size',
                'clang-32-O1-findutils-find.pkl#debug_strftime',
                ...
            ]
}

Examples of split_func_path file for callers:

splitFuncDict={
    'train':[
                'clang-32-O3-utillinux-dmesg.pkl#strnchr#print_record#283',
                'gcc-32-O3-coreutils-numfmt.pkl#process_line#main#386',
                'gcc-32-O3-coreutils-numfmt.pkl#process_line#main#557',
                ...
            ],
    'test': [
                'clang-32-O0-utillinux-utmpdump.pkl#gettok#undump#123',
                'gcc-32-O0-utillinux-ionice.pkl#ioprio_print#main#315',
                'clang-32-O1-inetutils-ping.pkl#ping_set_packetsize#ping_echo#24',
                ...
            ]
}

Testing RNN Model

Usage:

python eval.py [options] -d data_folder -f split_func_path -e embed_path -m model_dir -o output_dir

Link to eval.py

data_folder: The folder saves the binary information.
split_func_path: The file saves the training & testing function names.
embed_path: The file is the output file of save_embeddings.py, which saves the embedding vector of each instruction.
model_dir: The directory is used to save the trained model & log information.
output_dir: The directory is used to saved the predicted results and true labels of each function for each model.

Options:

-t: Type of output labels. Possible value: num_args, type#0, type#1, ... (Default value: num_args)
- num_args: The trained model is used to predict the number of arguments for each function.
- type#0: The trained model is used to predict the type of first argument for each function.
- type#1: The trained model is used to predict the type of second argument for each function.
- ...
-dt: Type of input data. Possible value: caller and callee. (Default value: callee)
- caller: The input data is from caller.
- callee: The input data is function body.
-pn: Number of Processes. (Default value: 40)
-ed: Dimension of embedding vector for each instruction. (Default value: 256)
-ml: Maximum length of input sequences. (Default value: 500)
-nc: Number of Classes. (Default value: 16)
-do: Dropout value.(Default value: 1.0)
-nl: Number of layers in RNN. (Default value: 3)
-b: Batch size. (Default value: 256)

Disclaimer

The code is research-quality proof of concept, and is still under development for more features and bug-fixing.

References

Neural Nets Can Learn Function Type Signatures From Binaries

Zheng Leong Chua, Shiqi Shen, Prateek Saxena, Zhenkai Liang.

In the 26th USENIX Security Symposium (Usenix Security 2017)

Project Members

Zheng Leong Chua, Shiqi Shen, Prateek Saxena, Zhenkai Liang, Valentin Ghita.

shensq04 / EKLAVYA

readme

Neural Nets Can Learn Function Type Signatures From Binaries

Authors

Dataset

Function Representation

Binary Representation

Code

Requirements

Train Embedding Model

Prepare the input file for training embedding model

Train the embedding model

Save the embedding vector

Train RNN Model

split_func_path Format

Testing RNN Model

Disclaimer

References

Project Members