Vulnerability Detection with Adaptive Fusion of Static and Dynamic Features

Binary vulnerability detection is critical in software security. The existing methods mainly rely on the static features of binary code, ignoring the importance of dynamic features, which limits the performance of the detection model. Therefore, an adaptive fusion method combining static and dynamic features is proposed. Firstly, through statistical analysis of binary code data, a new loss function is designed to capture the relationship between static and dynamic features. Then, the Long Short-Term Memory (LSTM) network was used to optimize feature selection. Then, a fusion mechanism based on a gated recurrent neural network was constructed, and weights were assigned to different features to enhance feature representation. Finally, the cosine similarity value of the known vulnerability function and the function to be tested is calculated to detect the potential vulnerability function. The experimental results show that the method improves by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1.7 \%-7.8 \%$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1 \%-5 \%$</tex> on ACC and AUC and by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$15.5 \%-33 \%$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$14.3 \%-29 \%$</tex> on Rank1 and MRR10, respectively.