Feature-Focused Defense: The Role of Attention Mechanisms in Enhancing Adversarial Attack Detection

Ryada, Mahesh (2025) Feature-Focused Defense: The Role of Attention Mechanisms in Enhancing Adversarial Attack Detection. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (760kB) | Preview
Preview	PDF (Configuration Manual) Download (553kB) | Preview

Abstract

Adversarial manipulation of sensor data poses a major challenge in safety-critical industrial environments, where even subtle perturbations can disrupt predictive maintenance models and lead to unreliable engine health assessment. Traditional machine learning and deep learning approaches used in such applications remain vulnerable to these perturbations, as their fixed feature-weighting strategies and static representations struggle to capture the nuanced deviations introduced by adversarial interference. Existing defense` strategies—including adversarial training, feature-space regularization, and domain adaptation—show performance gains but often fail to generalize across varying attack scenarios, leading to reduced robustness in dynamic sensor conditions. This study investigates a feature-tokenized transformer architecture that leverages attention-based feature interaction modelling to strengthen adversarial detection in turbofan engine sensor data. The model dynamically recalculates feature relevance, enabling improved identification of subtle inconsistencies in manipulated inputs when compared with conventional architectures. A comprehensive evaluation across multiple baseline machine learning and deep learning models demonstrates that the transformer-based approach achieves the strongest performance, recording an accuracy of 0.93 with substantial improvements in precision (0.95), and F1-score (0.87). The feature-level attention mechanism contributes to stable decision boundaries even under perturbed conditions, enabling significantly better discrimination between clean and adversarial samples. These findings highlight the suitability of attention-driven architectures for adversarial resilience in sensor-driven predictive maintenance and point toward their potential for enhancing the reliability of industrial monitoring systems exposed to manipulation risks.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Qayum, Abdul UNSPECIFIED
Uncontrolled Keywords:	Adversarial Attack Detection; Attention Mechanisms; FT-Transformer; Predictive Maintenance
Subjects:	Q Science > QA Mathematics > Computer software > Computer Security T Technology > T Technology (General) > Information Technology > Computer software > Computer Security Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	09 Sep 2026 08:26
Last Modified:	09 Sep 2026 08:26
URI:	https://norma.ncirl.ie/id/eprint/9903

Actions (login required)

View Item