Investigation on reinforcement learning for dynamic firewall configuration and its effect on edge devices

Sammanfattning

Cybersecurity has never been more relevant than it is today, as attacks become increasingly sophisticated. To keep devices safe from threats, a new type of smart firewall has become an active area of research. This type of firewall can dynami- cally configure itself based on incoming network traffic using reinforcement learn- ing. This research has shown promising results; however, its impact on resource usage on edge devices with limited hardware resources has not been properly in- vestigated. In this thesis, we will investigate how different reinforcement learning models (DQN, Double DQN, Dueling DQN, and 3DQN) and levels of complexity affect the performance of smart firewalls when deployed on edge devices. The re- sults indicate that models with 10k-50k parameters achieve the optimal trade-off between F1 score, memory usage, and inference time.