# Deep Reinforcement Learning for Cyber Security

**Authors:** Thanh Thi Nguyen, Vijay Janapa Reddi  
**arXiv:** [1906.05799](https://arxiv.org/abs/1906.05799)  
**Date:** 2019-06-13  
**Summary page:** https://www.summarizepaper.com/en/paper/1906.05799v4/

## Key points

- Internet-connected systems are more vulnerable to cyber attacks than ever before
- Protecting mechanisms need to be responsive, adaptive, and scalable
- Machine learning methods, particularly deep reinforcement learning (DRL), have been proposed as a solution for these issues
- DRL is highly capable of solving complex, dynamic, and high-dimensional cyber defense problems
- The paper presents a comprehensive survey of DRL approaches developed for cyber security
- The survey covers various aspects such as DRL-based security methods for cyber-physical systems, autonomous intrusion detection techniques, and multiagent DRL-based game theory simulations for defense strategies against cyber attacks
- Potential topics where DRL has not yet been applied in cyber security are also discussed
- Extensive discussions and future research directions on DRL-based cyber security are provided
- There is great potential to further research this area to develop more effective solutions that can adapt to the stochastic environment posed by modern cybersecurity threats.

## Summary

The increasing scale of Internet-connected systems has made them more vulnerable to cyber attacks than ever before. To address the complexity and dynamics of these attacks, protecting mechanisms need to be responsive, adaptive, and scalable. Machine learning methods, particularly deep reinforcement learning (DRL), have been widely proposed as a solution to these issues. By incorporating deep learning into traditional RL, DRL is highly capable of solving complex, dynamic, and high-dimensional cyber defense problems. This paper presents a comprehensive survey of DRL approaches developed for cyber security. The structure of the survey is presented in Fig. 4, which includes different subsections covering various aspects such as DRL-based security methods for cyber-physical systems, autonomous intrusion detection techniques, and multiagent DRL-based game theory simulations for defense strategies against cyber attacks. The survey also touches on potential topics where DRL has not yet been applied in cyber security such as combining host-based and network-based intrusion detection systems or investigating methods that can deal with unknown cyber attacks. Additionally, extensive discussions and future research directions on DRL-based cyber security are provided. Overall, this review provides the foundations for exploring the potential of emerging DRL to cope with increasingly complex cyber security problems. As highlighted in the IEEE Transactions on Neural Networks and Learning Systems article (DOI: 10.1109/TNNLS.2021.3121870), early access edition, there is great potential to further research this area to develop more effective solutions that can adapt to the stochastic environment posed by modern cybersecurity threats.