Research

Publications

Abstract: This paper presents a novel approach to defending against packet-mutation based adversarial attacks in Network Intrusion Detection Systems (NIDS). We introduce two key innovations: a genetic algorithm-based packet mutation technique for generating adversarial examples, and an adaptive defense mechanism that combines ensemble learning with meta-learning capabilities. Our defense system employs multiple base models (Random Forests, SVMs, LSTM networks, and CNNs) alongside a meta-learner that continuously adapts to evolving threats. Experimental results using the UNSW-NB15 dataset demonstrate that our adaptive defense system achieves a 93.7% detection rate, significantly outperforming traditional NIDS approaches which achieve only 52.3% detection rate. The system maintains high performance even under heavy network loads (up to 250,000 packets per second) with sub-second adaptation times. Notable improvements include a 43.4% increase in zero-day attack detection and a 39.4% improvement in detecting packet mutation attacks compared to traditional systems. The integration of meta-learning with ensemble methods proves particularly effective in addressing evolving attack patterns while maintaining low computational overhead.

Abstract: Spatial crowdsourcing (SC) platforms depend on reputation systems to ensure trust and quality in the execution of tasks. However, traditional centralised reputation mechanisms face scalability, privacy, and connectivity challenges. This paper presents a decentralised reputation tracking framework that utilises Tiny Machine Learning (TinyML). In this framework, lightweight machine learning models embedded in edge devices compute and maintain local reputation scores based on the outcomes of tasks. These local scores are then shared with a central task allocator, allowing for trust-aware assignments without exposing raw behavioural data. The proposed modular architecture facilitates on-device feature extraction and reputation updates, significantly reducing communication overhead. Experimental results using Yelp datasets indicate that our approach achieves competitive trust estimation accuracy, decreases communication by 45%, and a 4x reduction in latency compared to centralised baselines. This work establishes a foundation for privacy-preserving, scalable, and intelligent trust management in the next generation of SC systems.

Abstract: Network Intrusion Detection Systems (NIDS) are essential for cybersecurity monitoring but generate overwhelming volumes of alerts, with most having low importance. This paper introduces Hybrid-NIDS-AL, a novel framework that combines transfer learning with a hierarchical ensemble of active learning strategies for efficient NIDS alert classification. Unlike previous approaches that require extensive labeled datasets or rely solely on uncertainty sampling, our methodology incorporates adaptive stratified sampling across different alert partitions and leverages knowledge transfer from generic security alerts to environment-specific contexts. Experiments on a comprehensive dataset of 1.2 million alerts from a university SOC demonstrate that Hybrid-NIDS-AL achieves an F1-score of 0.973, outperforming traditional supervised approaches and state-of-the-art active learning methods. Our framework reaches 90% of peak performance with only 300 labeled samples-40% fewer than comparable methods. Additionally, Hybrid-NIDS-AL demonstrates superior resilience to concept drift, maintaining consistent performance over time. These results suggest significant potential for reducing human analyst workload while improving alert classification accuracy in operational security environments.

Abstract: Network Intrusion Detection Systems (NIDS) play a vital role in safeguarding Internet of Things (IoT) environments against escalating security threats. As attack vectors become increasingly sophisticated, traditional detection methods struggle to maintain effectiveness, especially when managing the vast quantities of heterogeneous data generated by IoT networks. This research introduces a novel Hybrid Ensemble Classification (HEC) framework that integrates advanced machine learning and deep learning techniques to enhance intrusion detection performance in IoT security. Our approach addresses key challenges through three primary innovations: (1) a Weighted Correlation Analysis technique that dynamically adapts feature selection thresholds based on statistical properties of datasets; (2) an enhanced Synthetic Minority Over-sampling Technique with Entropy-based Neighborhood Calibration to address class imbalance issues; and (3) a hierarchical ensemble architecture that combines temporal pattern analysis via Gated Recurrent Units with Attention and feature space optimization through an enhanced Random Forest algorithm. Comprehensive evaluations across UNSW-NB15, CIC-IDS2018, and IoTID20 datasets demonstrate that our proposed framework consistently outperforms state-of-the-art approaches, achieving accuracy improvements of up to 3.91% and F1-score gains of up to 4.16%, while maintaining computational efficiency suitable for resource-constrained IoT environments.

Abstract: This paper presents a novel approach to defending against packet-mutation based adversarial attacks in Network Intrusion Detection Systems (NIDS). We introduce two key innovations: a genetic algorithm-based packet mutation technique for generating adversarial examples, and an adaptive defense mechanism that combines ensemble learning with meta-learning capabilities. Our defense system employs multiple base models (Random Forests, SVMs, LSTM networks, and CNNs) alongside a meta-learner that continuously adapts to evolving threats. Experimental results using the UNSW-NB15 dataset demonstrate that our adaptive defense system achieves a 93.7% detection rate, significantly outperforming traditional NIDS approaches which achieve only 52.3% detection rate. The system maintains high performance even under heavy network loads (up to 250,000 packets per second) with sub-second adaptation times. Notable improvements include a 43.4% increase in zero-day attack detection and a 39.4% improvement in detecting packet mutation attacks compared to traditional systems. The integration of meta-learning with ensemble methods proves particularly effective in addressing evolving attack patterns while maintaining low computational overhead.

Current Internship Projects

This research explores how MetaHuman-based digital agents can serve as a human-centred, adaptive feedback interface within higher education Learning Management Systems (LMS). The goal is to move beyond correctness-only automated feedback by enabling conversational, dialogic interactions that help students understand their performance and take meaningful next steps.

Why this matters
Constructive and timely feedback is critical for learning, yet providing personalised feedback at scale remains a challenge—especially in large or online classes. This project proposes a Digital Human interface that mediates learning analytics through natural conversation, delivering feedback that is:

• Personalised to the learner’s needs

• Actionable with clear guidance on what to do next

• Reflective to support self-regulated learning and improvement

Pedagogical Approach
From a teaching and learning perspective, the MetaHuman agent will support scaffolded guidance by:

• Contextualising assessment outcomes in plain language

• Highlighting student strengths and areas for improvement

• Offering structured reflection prompts

• Recommending specific learning resources and next steps

Privacy-by-Design & Security
Trust is central to this research. The system will be designed to protect student privacy and minimise risk through:

• Operating on derived analytics, not raw student data

• Using pseudonymous identifiers where possible

• Providing learners with meaningful control over personalisation and interaction settings

• Applying secure integration using least-privilege access, standard interfaces, and strong authentication, authorisation, and audit logging

Expected Impact
By combining human-centred interaction through Digital Humans with privacy-preserving and secure system architecture, this research addresses a key gap in developing trustworthy adaptive feedback systems for higher education at scale.