Deep Learning-Based Hierarchical Attention Mechanism for Risk Prediction

Vishal Sahebrao Badgujar, Chandrashekhar M. Raut

Abstract: Cyber-attacks on information systems, which have become extremely technical and complicated, have changed the normal instruments of risk assessment and mitigation, thus lessening their efficiency and usability. This led to the development of advanced approaches that are more accurate in predicting the vulnerabilities of the system and the potential impact of cyber-attacks. Basically, most of these new models come from the old one-dimensional analytical methods which are unable to catch the whole picture, so they do not have the depth and device that will make them flexible enough to be able to take in a variety of changes in the interaction of networks and the behavior of the different network components. These negative points show the need for a complex, flexible analysis framework. We introduce an innovative deep learning method that uses a Hierarchical Attention Mechanism to predict risk. The strategy is more accurate in risk prediction compared to classical models that are way worse. This mechanism applies a two-tiered weighing system to different network parameters, providing a clearer picture of security risks, and their kinds that can be observed. As per our approach, the key to becoming more competitive in the domain of advanced cybersecurity and defeating the highly sophisticated modern cyber-attacks is the use of dynamic and multidimensional security analytics. Cybersecurity was traditionally built based on quantifiable security issues, individual experiences, and attack event alarms. The incursion of infrastructure defenses, however, remains potent, as exhibited by cyber adversaries’ advanced capabilities. This method fuses all the information and assigns a single vector concerned with the overall risk exposed by the attack. From the conducted experiments, Access of Abnormal Port has the highest attention score of 0.85 with respect to other attributes, demonstrating the possible threats within various circumstances. The Hierarchical Attention Mechanism employed in the entire attack prediction and assessment provided a remarkable accuracy of 97.3% in detection of attacks and F1-score of 98%.

External IDs:doi:10.1007/978-981-96-4142-0_37