Enhanced Host-Based Intrusion Detection Using System Call Traces

Abstract

To detect zero-day attacks in modern systems, several host-based intrusion detection systems are proposed using the newly compiled ADFA-LD dataset. These techniques use the system call traces of the dataset to detect anomalies, but generally they suffer either from high computational cost as in window-based techniques or low detection rate as in frequency-based techniques. To enhance the accuracy and speed, we propose a host-based intrusion detection system based on distinct short sequences extraction from traces of system calls with a novel algorithm to detect anomalies. To the best of our knowledge, the obtained results of the proposed system are superior to all up-to-date published systems in terms of computational cost and learning time. The obtained detection rate is also much higher than almost all compared systems and is very close to the highest result. In particular, the proposed system provides the best combination of high detection rate and very small learning time. The developed prototype achieved 90.48% detection rate, 22.5% false alarm rate, and a learning time of about 30 seconds. This provides high capability to detect zero-day attacks and also makes it flexible to cope with any environmental changes since it can learn quickly and incrementally without the need to rebuild the whole classifier from scratch.

Combining dynamic and static host intrusion detection features using variational long short-term memory recurrent autoencoder

“…The results showed an average accuracy rate of 70 %, with a FPR of 20 %. The authors [14] propose an algorithm to extract distinct short sequences from traces of system calls. Features after extracting are feed to the machine learning model such as SVM, kNN to detect anomalies.…”