PBDT: Python Backdoor Detection Model Based on Combined Features

Table 6

Performance comparison of different feature combinations.

	Features	Accuracy	Precision	Recall (TPR)	TNR	F1-score
#1	Call	0.9342	0.8805	0.8696	0.9575	0.8750
#2	Text	0.9391	0.9026	0.8634	0.9664	0.8825
#3	Opcode	0.9441	0.8757	0.9193	0.9530	0.8970
	#1+#2	0.9589	0.9146	0.9317	0.9687	0.9231
	#1+#3	0.9638	0.9264	0.9379	0.9732	0.9321
	#2+#3	0.9473	0.8817	0.9255	0.9553	0.9030
PBDT	#1+#2+#3	0.9770	0.9623	0.9503	0.9866	0.9563