Binary Code Vulnerability Detection Based on Multi-level Feature Fusion

The existence of software vulnerabilities will cause serious network attacks and information leakage problems. Timely accurate detection in has become a research focus on the security field. Most existing work only considers instruction-level features, which to some extent overlooks certain syntax semantic assembly code segments, affecting accuracy model. In this paper, we propose binary vulnerability model based multi-level feature fusion. both word-level features features. order solve problem that traditional text embedding methods cannot handle polysemy, paper uses Embeddings from Language Models (ELMo) obtain dynamic word vectors containing semantics other information. Considering grammatical structure segment, randomly embeds normalized segment represent it. Then bidirectional Gated Recurrent Unit (GRU) extract sequence respectively. Then, weighted fusion method is used study impact different performance. During training, adding standard deviation regularization constrain parameters can prevent occurrence overfitting To evaluate our proposed method, conduct experiments two datasets. Our achieves an F1-score 98.9 percent Juliet Test Suite dataset 87.7 NDSS18 (Whole) dataset. experimental results show improve detection.