An Angular-Margin-Based Binary Function Contrastive Learning Framework
Existing code similarity detection models primarily focus on constructing encoders,with lim-ited research on loss functions in deep learning.To address the overlooked issue of evaluating embedded binary function vectors,this paper proposes an angular-margin-based binary code contrastive learning framework(AngCLF).By optimizing the objective function of contrastive learning,the model's accuracy and convergence speed are enhanced.Besides,the study analyzes the reasons for the model's effectiveness and introduces multiple metrics for evaluating binary code vector spaces.The experimental results validate the accuracy of the AngCLF.The AngCLF surpasses six models including the jTrans model in accuracy,and has faster convergence speed and obvious advantages in alignment and uniformity metrics.
NETL
NSTL
万方数据
