Advanced Machine Learning for Privacy-Preserving Intrusion Detection in IoT Networks

Abstract

The Internet of Things (IoT) is a groundbreaking technology that integrates smart devices into communication networks, reducing human intervention and fueling innovation across various applications, from smart cities to intelligent transportation. However, its heterogeneous and highly interconnected nature introduces significant vulnerabilities, expanding attack surfaces that are increasingly targeted by sophisticated cyber threats. To address this, machine learning (ML), especially deep learning (DL) has shown the potential to equip IoT components with intelligent cyberattack detection modules, enabling IoT components to detect and respond to cyberattacks by analyzing vast data streams. However, traditional approaches that transmit vast amounts of user data to centralized models for processing pose significant data privacy risks, exposing sensitive information to potential misuse by the model’s owner. This highlights the need for advanced distributed ML/DL methods that can accurately detect cyberattacks and preserve user data privacy.

Additionally, IoT network data often contains highly sensitive user information (e.g., location details and biometric data) making it susceptible to privacy breaches when being analyzed for cyberattack detection using ML models. To tackle this problem, homomorphic encryption (HE) offers a powerful technique to integrate with ML/DL models, enabling innovative privacy-preserving ML (PPML) approaches that safeguard the confidentiality of user data. By employing this cryptographic technique, ML/DL models can compute directly on ciphertext without the need for decryption. However, PPML using HE is still an emerging field, and designing HE-based learning algorithms remains a significant challenge due to the high computational overhead of HE and the complexity of ML/DL models. Therefore, the design of HE-based PPML algorithms to balance security, accuracy, and performance is crucial for practical deployment in real-world IoT systems.

In this thesis, we contribute to addressing both security and privacy concerns in two emerging IoT ecosystems: (1) developing a novel framework that integrates federated learning (FL) and HE for privacy-preserving intrusion detection in resource-constrained Internet of Vehicle (IoV) networks, and (2) designing a privacy-preserving cyberattack detection in blockchain-based IoT networks using our proposed deep neural network (DNN) training algorithm for HE-encrypted data and our novel privacy-preserving distributed-cloud native learning algorithm. Following the Single-Instruction-Multiple-Data (SIMD) manner, our proposed packing algorithms enable an efficient privacy-preserving learning/inference for encrypted data, achieving results nearly identical to the non-encrypted approach, approximately from 0.01 to 0.8%. These contributions demonstrate the effectiveness and potential of our research in leveraging advanced ML to tackle security challenges and ensure data confidentiality within real-world IoT ecosystems.