Adversarial Machine Learning: Difficulties in Applying Machine Learning to Existing Cybersecurity Systems

Abstract

Machine learning is an attractive tool to make use of in various areas of computer science. It allows us to take a hands-off approach in various situations where previously manual work was required. One such area machine learning has not yet been applied entirely successfully is cybersecurity. The issue here is that most classical machine learning models do not consider the possibility of an adversary purposely attempting to mislead the machine learning system. If the possibility that incoming data will be deliberately crafted to mislead and break the machine learning system, these systems are useless in a cybersecurity setting. Taking this into account may allow us to modify existing security systems and introduce the power of machine learning to them.