Sheikh Muhammad Farjad

On-device deep learning (DL) is reshaping structural health monitoring (SHM), but deploying models at the edge makes them increasingly vulnerable to adversarial attacks, particularly model extraction. This threat is severe for critical infrastructure, given the limited security focus of Edge AI systems. This dissertation develops a secure Edge AI system for SHM through four key contributions: (a) an adversarial risk taxonomy for assessing SHM system susceptibility to adversarial attacks, (b) a secure edge architecture and testbed for deployment, (c) an investigation of model extraction attacks using eBPF-collected metrics, and (d) a novel kernel-level defense to secure deployment.

The deployment of deep learning (DL) models on edge devices offers significant opportunities for structural health monitoring (SHM), particularly by enabling localized, low-latency inference in built environments. However, practical deployment on resource-constrained platforms faces barriers in scalability, reliability, and security. In response, we introduce Zeki, a unified, security-aware, containerized pipeline for deploying DL models for edge-based SHM. Zeki unifies model optimization (via LiteRT and quantization), container-level hardening, and benchmarking-driven model-device co-selection into a reproducible workflow. We evaluate Zeki by deploying convolutional neural networks (CNNs) for crack detection on Raspberry Pi 4 and BeagleBone AI-64 along with a server. Results show significant improvements in inference latency and memory efficiency compared to unoptimized baselines. Beyond performance, Zeki establishes a systematic methodology for safe, resilient, and evidence-based edge deployment of DL in safety-critical SHM settings, enabling long-term infrastructure monitoring.

Structural Health Monitoring (SHM) plays a crucial role in maintaining aging and critical infrastructure, supporting applications such as smart cities and digital twinning. These applications demand machine learning models capable of processing large volumes of real-time sensor data at the network edge. However, existing approaches often neglect the challenges of deploying machine learning models at the edge or are constrained by vendor-specific platforms. This paper introduces a scalable and secure edge-computing reference architecture tailored for data-driven SHM. We share practical insights from deploying this architecture at the Memorial Bridge in New Hampshire, US, referred to as the Living Bridge project. Our solution integrates a commercial data acquisition system with off-the-shelf hardware running an open-source edge-computing platform, remotely managed and scaled through cloud services. To support the development of data-driven SHM systems, we propose a resource consumption benchmarking framework called edgeOps to evaluate the performance of machine learning models on edge devices. We study this framework by collecting resource utilization data for machine learning models typically used in SHM applications on two different edge computing hardware platforms. edgeOps was specifically studied on off-the-shelf Linux and ARM-based edge devices. Our findings demonstrate the impact of platform and model selection on system performance, providing actionable guidance for edge-based SHM system design.