Detecting Malware with an Ensemble Method Based on Deep Neural Network

Table 4

The experiment results for different LSTM networks.

Strategy	Models	Accuracy (%)	AUC	TPR (FPR = 0.1%)	EER (%)	Training times (h)
TAP	LSTM ()	71.43	0.8863	52.03	-	0.41
	LSTM (α = 60)	87.13	0.9791	81.67	6.45	0.54
	LSTM (α = 80)	91.56	0.9854	85.39	4.88	0.89
	LSTM (α = 120)	94.86	0.9931	91.53	3.17	-
Truncated BPTT	LSTM (α = 30)	94.37	0.9928	91.11	3.10	-
	LSTM (α = 60)	96.83	0.9950	93.37	-	-
	LSTM (α = 80)	98.08	0.9989	95.13	-	1.24
	LSTM (α = 120)	98.47	0.9993	96.81	-	1.36
	LSTM (α = 180)	97.82	0.9987	95.42	-	1.53
Truncated BPTT + subsequence selection	LSTM (α = 120, = 90%)	97.98	0.9988	95.01	1.34	1.16
	LSTM (α = 120, = 95%)	98.83	0.9997	97.22	0.84	-
	LSTM (α = 120, = 99%)	98.66	0.9996	96.69	0.92	-
	LSTM (α = 120, = 97.5%)	99.13	0.9999	98.69	0.54	1.34