IOT threats detection using few shots learning
Existing IoT threat detection methods lack robustness due to the diverse array of potential attack vectors. Currently, most methods are trained and tested using simulated datasets and do not perform well with unseen samples in real-world applications. In this project, we propose a novel few short...
| Main Author: | |
|---|---|
| Format: | Final Year Project / Dissertation / Thesis |
| Published: |
2024
|
| Subjects: | |
| Online Access: | http://eprints.utar.edu.my/6634/ http://eprints.utar.edu.my/6634/1/fyp_CS_2024_CCH.pdf |
| Summary: | Existing IoT threat detection methods lack robustness due to the diverse array of potential
attack vectors. Currently, most methods are trained and tested using simulated datasets and do
not perform well with unseen samples in real-world applications. In this project, we propose a
novel few short learning leveraging Large Language Models (LLMs) to improve model
robustness in IoT threat detection. Firstly, we develop two specialized LLM models: a text
classification model based on DistilBERT and the few shots learning model using Sentence
Transformer Fine-Tuning model (SetFit) framework. The DistilBERT threats detection model
method performed well with an accuracy of 99.998% due to better semantics and contextual
understanding as compared to existing flow statistical analysis. The few-shot learning model
demonstrated remarkable performance with an accuracy of 0.89%, despite being trained on a
limited amount of data. For unseen samples, we designed a few-shot retraining (FSR)
methodology to adapt and learn new attack vectors across multiple variants using transfer
learning. The experimental results showed a 90% improvement in accuracy on unseen threats
when implemented in a real-world NIDS. |
|---|