Binary Code Similarity Detection with Multi-Similarity Loss

This is the source code of our solutions in CCF BDCI 2022 competition held on DataFountain. We introduce advanced miner and loss algorithms from Deep Metric Learning into Binary Code Similarity Detection.

Environments

Python                    3.10.4
pytorch                   1.12.1 (py3.10_cuda10.2_cudnn7.6.5_0)
pytorch-lightning         1.7.3 
torch-geometric(PyG)      2.1.0
pytorch-metric-learning   1.5.2
scikit-learn              1.1.1 

Run

bash train.sh && bash test.sh

Results

Our model records are in model/lstm_gatedgcn-e/version_0.

Model	Test MAP
Norm weighted miner + Triplet loss	0.855586
MS miner + Triplet loss	0.883205
MS miner + MS loss (Ours)	0.922685

Reference

• Multi-Similarity Loss with General Pair Weighting for Deep Metric Learning
• CodeCMR: Cross-Modal Retrieval For Function-Level Binary Source Code Matching