8002824268

18 Bartosz Bielski, Przemysław Klęsk

Table 1. Operating systems classes

Pos.	OS class	Number of examples
1.	Windows	599
2.	linux	349
3.	bsd	169
4.	other	164
5.	Solaris	132
6.	unix	94
7.	mac_os	89
8.	network modem/router	83
9.	cisco device	44
10.	network printer	39

	MLP(2 layers)	MLP(binearized)	RBF (Spread - 0,1)	MLP (3 layers)
■ Learning set	99,46%	81,30%	100%	99,80%
■ Testingset	88%	67%	95,75%	97,70%

Figurę 3. Results of recognition by many kinds of neural networks

When the knowledge and trust drops and especially when we cannot trust remote systems are clean and unmodified, neural network get really good prediction results. The good scores are especially valuable when there is no strict signature in the database of operating systems so that system cannot be recognised at all.

As seen on Figurę 4 we compared best neural network from earlier researches (3-layer multi layer perceptron network - MLP with 10 neurons in every hidden layer) with multi-stage neural network architecture that can detect operating system version