On the Effectiveness of Machine and Deep Learning for Cyber Security

Cybersecurity concerns, discussions, and technology have been active since the inception of the Internet. While at the same time, computation power escalated with the highly improved GPUs, enabling machine learning to be a reality. Only recently, however, the application aspect of machine learning and deep learning has become more widespread and incorporated into industry. The interconnection of deep learning and cybersecurity poses very interesting questions and concerns. Traditional cyberattack tactics followed rule based algorithms that are modulated by certain fixed sets of criteria. Now however, with deep learning, not only are these criteria self-learned, the algorithms are trained in such a way that they continue to improve over time in completely unpredictable ways. This is most commonly seen in the cyber realm for detection of unauthorized activity, fraudulent behavior, and malicious foreign servers to name a few. This article analyses at a technical depth the power and performance of these systems. Ultimately, the conclusion of this article claims that the current level of model sophistication is not nearly high enough for effective adoption of these systems. Unfortunately these models still have certain shortcomings such as the ability to be influenced by adversarial attacks, and need constant re-training to perform over time.