A semi-self-taught network intrusion detection system

The ever increasing threat and complexity of modern cyber-attacks requires search for integrated and flexible intelligent defense mechanisms. Such approaches can provide optimal countermeasures, reliable credentials extraction and self-adjusting potential. Given the widespread scale of modern networks and the complexity of cyber-attacks, the problem of self-adaptation goes far beyond the capabilities of network Intrusion Detection Systems (IDS). The main weakness of IDS is the fact that they cannot adapt to new network conditions (zero day attacks). This research tries to overcome the above limitation, by introducing a Semi-supervised Discriminant Autoencoder (AUE) which combines Denoising AUEs with a heuristic method of class separation. In essence, the proposed algorithm learns to remodel the displaced specimens instead of the original ones in the super-sphere defined by their closest neighbors. The purpose is to understand the nature of an attack, based on generalized transformed features derived directly from unknown web environments and data.