CodeCMR: Cross-Modal Retrieval For Function-Level Binary Source Code Matching (NeurIPS-2020)

• pandas=0.25.1
• networkx=2.3

Trainset: 30,000 Validset: 10,000 Testset: 10,000

Each dataset has 33 columns, the first column is the source code, the other columns are the corresponding binary code on 32 combinations of different compilers (gcc/clang), different platforms (x86/x64/arm/arm64) and different optimizations (O0/O1/O2/O3). Please first download the data from google cloud and uncompress it:

7z x all-arch-nx.zip

import pandas as pd
import networkx as nx

df = pd.read_pickle('test.pkl')
print(df.columns)                   # 33 columns

sample = df.iloc[0]
src, bin = sample['c_label'], sample['gcc-x64-O0']

print(src)                          # character-level source code
g = nx.read_gpickle(bin)
print(g.graph)                      # binary code literal features, we only use c_int and c_str
print(g.nodes.data('feat'))         # binary code CFG features