Empirical Evaluation of Noise Influence on Supervised Machine Learning Algorithms Using Intrusion Detection Datasets

Table 16

The results of running ML algorithms on an injected and then filtered intrusion dataset.

Noise (%)	ML algorithm	NSL-KDD			UNSW-NB15
		Accuracy	Precision	Recall	Accuracy	Precision	Recall
Baseline	DT (J48)	81.5339	0.858	0.815	100	1.000	1.000
	RF	80.4516	0.852	0.805	98.4903	0.985	0.985
	SVM	75.3948	0.802	0.754	100	1.000	1.000
	ANN	77.7147	0.817	0.777	100	1.000	1.000
	NB	76.1178	0.809	0.761	87.4435	0.884	0.874
5% noise injected then filtering	DT (J48)	77.9026	0.828	0.779	100	1.000	1.000
	RF	77.8277	0.749	0.778	97.051	0.973	0.971
	SVM	77.6779	0.715	0.777	100	1.000	1.000
	ANN	77.8652	0.828	0.779	100	1.000	1.000
	NB	77.6779	0.715	0.777	74.2945	0.795	0.743
10% noise injected then filtering	DT (J48)	73.97	0.742	0.740	100	1.000	1.000
	RF	73.8577	0.682	0.739	95.0076	0.957	0.950
	SVM	73.8202	0.676	0.738	100	1.000	1.000
	ANN	74.0824	0.779	0.741	100	1.000	1.000
	NB	73.8202	0.676	0.738	74.2055	0.795	0.742
20% noise injected then filtering	DT (J48)	75.2434	0.762	0.752	73.392	0.849	0.734
	RF	66.7416	0.589	0.667	91.5565	0.929	0.916
	SVM	67.7154	0.620	0.677	100	1.000	1.000
	ANN	67.6404	0.539	0.676	100	1.000	1.000
	NB	67.7154	0.620	0.677	74.0753	0.794	0.741
30% noise injected then filtering	DT (J48)	66.6292	0.676	0.666	70.2269	0.788	0.702
	RF	61.985	0.595	0.620	88.6869	0.900	0.887
	SVM	61.7228	0.573	0.617	99.9873	1.000	1.000
	ANN	61.7978	0.621	0.618	99.8062	0.998	0.998
	NB	61.7228	0.573	0.617	74.088	0.794	0.741