VulDistilBERT: A CPS Vulnerability Severity Prediction Method Based on Distillation Model

Table 13

Comparison of results related to similar works.

Researchers	Feature extraction method	Classifiers	Accuracy (%)	Precision (%)	Recall (%)	F1 scores (%)
Khazaei et al. [25]	TF-IDF	Fuzzy system	88.37	—	—	—
Spanos et al. [26]	Document-term matrix	Decision tree	79.12	75.54	71.26	73.02
		Neural network	78.26	73.59	70.24	71.68
		SVM	79.53	78.49	68.21	71.50
Nakagawa et al. [27]	One-hot	CNN	72.50	—	—	—
Wang et al. [28]	Feature vector	PCA + XGBOOST	92.38	—	—	—
Han et al. [29]	Word embedding	1-Layer CNN	81.60	81.80	81.50	81.60
Liu et al. [30]	Text mining	XGBoost	87.30	—	—	—
		CNN	92.04	—	—	—
		LSTM	93.73	—	—	—
		TextRCNN	93.95	—	—	—
Our method	Fine-tuned DistilBERT with DA	Linear	90.64	91.92	91.92	91.83
	Fine-tuned DistilBERT with OS	Linear	92.80	88.32	87.26	87.07
	Fine-tuned DistilBERT with DA and OS	Linear	96.62	97.11	97.06	97.05