Fine-tuning Deep Learning Models for Intrusion Detection
In today's world, the Internet is available almost everywhere so is the possibility of cyber-attacks. Many security systems have been built to prevent these attacks like Firewalls, Network protection software, Intrusion Detection Systems (IDS), and Security Information and Event Management Systems. No matter how secure the system is, there may always be a chance of getting compromised against an unknown attack. When such scenarios happen, the industries tend to design a new IDS model from scratch considering that unknown attack. Designing an IDS from scratch can be time-consuming. As a result, deep transfer learning is proposed by the proposed study. This research focuses on how developing Transfer Learning based on fine-tuning can affect the overall results of an IDS. 2 base models using CNN and LSTM are implemented for the same source task. Each model is transferred to learn about the same target domain of DoS and Heartbleed attacks. The proposed methodology is evaluated for 15% training size. The experiments prove that the best result F1 of 92.2% is acquired using the LSTM transfer model which is at least 10% more than ID3, LinearSVM, LR, and normal LSTM algorithm.
Authors: Harsh Mandali, Charlie Obimbo
- Date of Conference: 7-9 December 2021
- DOI: 10.20533/ICITST.2021.0024
- ISBN: 978-1-913572-39-6
- Conference Location: Virtual (London, UK)