Combat Mobile Evasive Malware via Skip-Gram-Based Malware Detection

Table 5

Dataset used in scenario 2 contains 21 less malware families which was excluded for zero-day testing; this caused small increase in detection performance. RF bested other methods with 96.12% accuracy.

	Malware			Benign			Total accuracy
	Precision	Recall	F1 measure	Precision (%)	Recall (%)	F1 measure (%)
SVM	87.6	79.8	83.5	75.8	84.8	80	81.94%
Random forest	97.4	95.8	96.6	94.4	96.6	95.5	96.12%
Decision tree	92.5	93.9	93.2	91.6	89.7	90.6	92.04%
Random subspace	96.4	94.6	95.5	92.9	95.2	94	94.86%
SGD	70.4	97.4	81.7	92.7	45	60.6	75.04%
KNN	93.5	97.1	95.3	95.9	90.9	93.4	94.48%
Ensemble 1	94.8	96.9	95.8	95.7	92.8	94.2	95.15%