Browsing by Author "Cao, Hung"

Now showing 1 - 8 of 8
  • Thumbnail Image
    Item
    A privacy-aware fall detection system for aging-in-place environments
    (University of New Brunswick, 2025-08) Rahimi Azghadi, Seyed Alireza; Cao, Hung; Palma, Francis
    Falls are a major threat to the health and independence of older adults, making effective fall detection critical in smart healthcare systems. Traditional approaches face challenges like limited labeled data and privacy concerns from centralized data collection. This thesis introduces a privacy-preserving fall detection framework that integrates three key systems: (1) a semi-supervised federated learning model for wearable-based fall detection without requiring labeled data; (2) an adaptive indoor localization technique using a SLAM-enabled robot for autonomous WiFi and BLE fingerprinting; and (3) a multi-stage response system combining wearable alerts, robotic navigation, and vision-based verification. The Semi-supervised Federated Fall Detection (SF2D) model enables devices to learn collaboratively while safeguarding privacy. The robotic system builds a detailed radio map for precise localization, and the integrated system confirms falls through visual validation. Experimental results show improved detection accuracy, fewer false alarms, and enhanced privacy and resource consumption. This work presents a scalable, ethical solution to support aging-in-place through intelligent fall detection.
  • Thumbnail Image
    Item
    Developing a Resource-Constraint EdgeAI model for Surface Defect Detection
    (2023) Mih, Atah Nuh; Cao, Hung; Kawnine, Asfia; Wachowicz, Monica
    Resource constraints have restricted several EdgeAI applications to machine learning inference approaches, where models are trained on the cloud and deployed to the edge device. This poses challenges such as bandwidth, latency, and privacy associated with storing data off-site for model building. Training on the edge device can overcome these challenges by eliminating the need to transfer data to another device for storage and model development. On-device training also provides robustness to data variations as models can be retrained on newly acquired data to improve performance. We therefore propose a lightweight EdgeAI architecture modified from Xception, for on-device training in a resource-constraint edge environment. We evaluate our model on a PCB defect detection task and compare its performance against existing lightweight models - MobileNetV2, EfficientNetV2B0, and MobileViT-XXS. The results of our experiment show that our model has a remarkable performance with a test accuracy of 73.45% without pre-training. This is comparable to the test accuracy of non-pre-trained MobileViT-XXS (75.40%) and much better than other non-pre-trained models (MobileNetV2 - 50.05%, EfficientNetV2B0 - 54.30%). The test accuracy of our model without pre-training is comparable to pre-trained MobileNetV2 model - 75.45% and better than pre-trained EfficientNetV2B0 model - 58.10%. In terms of memory efficiency, our model performs better than EfficientNetV2B0 and MobileViT-XXS. We find that the resource efficiency of machine learning models does not solely depend on the number of parameters but also depends on architectural considerations. Our method can be applied to other resource-constraint applications while maintaining significant performance.
  • Thumbnail Image
    Item
    Developing an analytics everywhere framework for the Internet of Things in smart city applications
    (University of New Brunswick, 2019) Cao, Hung; Wachowicz, Monica
    Despite many efforts on developing protocols, architectures, and physical infrastructures for the Internet of Things (IoT), previous research has failed to fully provide automated analytical capabilities for exploring IoT data streams in a timely way. Mobility and co-location, coupled with unprecedented volumes of data streams generated by geo-distributed IoT devices, create many data challenges for extracting meaningful insights. This research work aims at exploring an edge-fog-cloud continuum to develop automated analytical tasks for not only providing higher-level intelligence from continuous IoT data streams but also generating long-term predictions from accumulated IoT data streams. Towards this end, a conceptual framework, called “Analytics Everywhere”, is proposed to integrate analytical capabilities according to their data life-cycles using different computational resources. Three main pillars of this framework are introduced: resource capability, analytical capability, and data life-cycle. First, resource capability consists of a network of distributed compute nodes that can handle automated analytical tasks either independently or in parallel, concurrently or in a distributed manner. Second, analytical capability orchestrates the execution of algorithms to perform streaming descriptive, diagnostic, and predictive analytics. Finally, data life-cycles are designed to manage both continuous and accumulated IoT data streams. The research outcomes from a smart parking and a smart transit scenario have confirmed that a single computational resource is not sufficient to support all analytical capabilities that are needed for IoT applications. Moreover, the implemented architecture relied on an edge-fog-cloud continuum and offered some empirical advantages: (1) on-demand and scalable storage; (2) seamlessly coordination of automated analytical tasks; (3) awareness of the geo-distribution and mobility of IoT devices; (4) latency-sensitive data life-cycles; and (5) resource contention mitigation.
  • Thumbnail Image
    Item
    Hybrid LSTM–RLS day-ahead net load demand forecasting for Barbados distribution feeders
    (University of New Brunswick, 2025-08) Emani, Sai Venkata Krishna Sri Harsha; Cardenas Barrera, Julian; Cao, Hung
    Forecasting net load demand in island grids is challenging due to limited grid interconnections and high reliance on RES. Rapid integration of RES on islands introduces forecasting challenges due to their intermittent and stochastic nature. This study develops and evaluates a hybrid deep learning algorithm for Barbados’s distribution feeders to accurately predict net load demand. Twelve feeders are selected, and their net load demand data is combined with calendar and weather features to develop a robust algorithm. It is trained on feature vectors comprising 24 hours of past and forecasted features. The architecture comprises of two LSTM networks that are trained to predict the absolute and change in net load demand values respectively. Their outputs are adaptively combined by an RLS combiner for the final forecast. Several experiments were conducted to evaluate the model’s performance by benchmarking it against other models.
  • Thumbnail Image
    Item
    Linguistic patterns and antipatterns detection and their impact on understandability and readability of APIs
    (University of New Brunswick, 2025-04) Dey, Krishno; Palma, Francis; Cao, Hung
    Application Programming Interfaces (APIs) allow distributed systems to expose their functionalities. Despite well-known API design rules and guidelines (patterns), many poor design practices (antipatterns) are prevalent in APIs. This thesis aims to (1) assess the linguistic design quality of APIs and (2) evaluate the impact of patterns and antipatterns on the understandability and readability of APIs through a survey of API developers. We rely on syntactic and semantic analyses to automatically assess the design quality of APIs. Syntactic analysis involves analyzing the structure and syntax of the APIs, while semantic analysis involves analyzing API documentation, descriptions, and parameters. We found that linguistic antipatterns are prevalent in APIs. Our detection algorithms achieve an average detection accuracy of 94%. The survey confirms that adherence to linguistic patterns significantly enhances the understandability and readability of APIs. Our findings will assist API developers in improving the design quality of their APIs.
  • Thumbnail Image
    Item
    Multilingual Phishing Email Detection Using Lightweight Federated Learning
    (IEEE, 2025-08-26) Staples, Dakota; Cao, Hung; Hakak, Saqib; Cook, Paul
    Given the escalating global threat of phishing emails, it is imperative to develop effective solutions to mitigate their potentially devastating impacts on society. This study endeavours to construct a federated multilingual spam detection system employing logistic regression, specifically targeting English, French, and Russian emails. This is the first work to the best of our knowledge which considers a non-deep learning setting for federated learning, and combines federated learning with multilingual phishing detection. Evaluation of the models is based on accuracy metrics which are compared with a most frequent class baseline. Our findings indicate that an optimal configuration comprises 10 clients undergoing 100 epochs of training with 100 rounds of federated learning, resulting in superior performance. Notably, this approach significantly outperforms the baseline, achieving an accuracy of 89.46% compared to 70%
  • Thumbnail Image
    Item
    Towards an advancement of federated learning framework: A unified approach for multi-tier spatial encoding, spatio-temporal modeling, and multi-global server architectures
    (University of New Brunswick, 2025-04) Kawnine, Asfia; Cao, Hung
    As machine learning (ML) applications continue to expand into distributed environments, ensuring privacy, efficiency, and scalability in model training has become a critical challenge. Federated Learning (FL) enables decentralized model training while preserving data privacy, but traditional frameworks struggle with spatial dependencies, real-time adaptation, and scalable aggregation. This research integrates three key advancements to enhance FL: Spatial Encoding and Multi-Tier Aggregation, Real-Time Spatio-Temporal Modeling, and Multi-Global Server Architectures. By integrating spatial encoding, FL improves prediction accuracy by an average of 6% in location-sensitive applications, while multi-tier aggregation addresses the geospatial relation. Spatio-temporal modeling combines spatial learning and temporal dependencies and reduces mean square error by 15% on average, enabling FL to adapt to real-time data streams in various applications. A multi-global server architecture enhances fault tolerance and system resilience. This unified FL framework is efficient, improving model accuracy and system stability in large-scale Artificial Intelligence (AI) applications.
  • Thumbnail Image
    Item
    Unlocking the benefits of transfer learning in edge-cloud computing environments
    (University of New Brunswick, 2024-04) Nuh Mih, Atah; Cao, Hung
    Transfer learning’s success motivates the need to understand its characteristics across cloud, edge, and edge-cloud computing paradigms. Thus, this extensive research evaluates the role of transfer learning in 1) cloud computing; 2) edge computing; and 3) edge-cloud computing. It first proposes a transfer learning approach to address the data limitation and model scalability challenges for machine learning in a cloud computing environment. Then, this study provides a model optimization for deep neural networks to improve hardware efficiency for training models on edge devices and investigates the role of transfer learning on resource consumption. Finally, a weight-averaging method is proposed for collaborative knowledge transfer across a unified edge and cloud computing environment to improve training performance for local edge models and global server models. The research conclusively shows that transfer learning benefits edge and cloud computing paradigms both individually and collaboratively.