Deep Neural Embedding for Software Vulnerability Discovery: Comparison and Optimization

Table 6

The number of vulnerable functions and nonvulnerable functions when fine tuning the parameters of CodeBERT. In the training set and verification set, aiming to make vulnerable functions account for 1/10 of the total number of functions, we added synthetic data to the original dataset.

Dataset	No of vul. Functions (real-world | SARD)	No. of total functions
Training set	8675 (1189 | 7486)	86759
Validation set	2891 (395 | 2495)	28919
Test set	399 (399 | 0)	26425