Fortifying IoT Devices: AI-Driven Intrusion Detection via Memory-Encoded Audio Signals

Document Type

Conference Proceeding

Publication Date

1-1-2023

Abstract

IoT devices have become an increasingly accessible target for evasive attacks, such as botnets, due to insecure network services, deprecated software components, unencrypted data communication, and other vulnerabilities. To address these security concerns, our work makes several significant contributions toward curating datasets and designing and developing a robust and effective Host-Based Intrusion Detection algorithm (HIDS) for IoT devices. The proposed algorithm leverages memory-based fingerprints to train a convolutional neural network (CNN) model. Our approach is based on the premise that despite the heterogeneity of IoT devices, the functionality of each IoT device is often unique and remains relatively constant throughout its lifespan. Thus, to develop an effective IDS algorithm based on anomaly detection, we encode the dynamic IoT device memory into sound wave signals, extract discriminable features such as MFCCs and Chroma features, and pass these deterministic fingerprints as feature vectors to a device-specific CNN model. When trained using three different testbed datasets, our model achieved 100% accuracy for known and anomaly memory instances. Additionally, the evaluation of our features and feature engineering process showed that our model, which is trained with the features from memory-encoded signals, is more reliable and robust than those algorithms that leverage raw memory bytes as feature vectors.

Publication Source (Journal or Book title)

Proceedings - 2023 IEEE Secure Development Conference, SecDev 2023

First Page

106

Last Page

117

This document is currently not available here.

Share

COinS