LBLinxue Bai

Linxue Bai

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Linxue Bai has a total of 23 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Secure and Enhanced Cyber-Threat Detection in IoMT Using Locally Deployed Large Language Models

    by Saadullah Farooq Abbasi, Muhammad Bilal, Xuefei Ding, Linxue Bai, Omid Pournik, Saif Ul Islam, Gregory Epiphaniou, Carsten Maple, Theodoros N. Arvanitis

    The Internet of Medical Things (IoMT) is growing exponentially with the advancements in image processing, large language models (LLMs), and other technologies. However, processing patient-identifiable information (PII) in cloud environments raises significant security and compliance concerns. Existing tools provide baseline threat detection; however, they miss complex vulnerabilities. Researchers used cloud-based LLMs such as GPT or Claude to enhance detection, however, these approaches compromise privacy. For this reason, this study proposes a locally deployed and cost-efficient LLM specifically, Llama 3.1 (8B parameters), to identify additional threats and recommend mitigation strategies. The locally deployed Llama 3.1 model identified 12 additional cyberthreats beyond SPYDERISK improving detection coverage and preserving privacy.

  • Open Access

    Membership Inference or Data Split Bias? Identifying False Positives in Synthetic Medical Image Privacy Audits

    by Linxue Bai, Omid Pournik, Xuefei Ding, Saadullah Farooq Abbasi, Le Zhang, Theodoros Arvanitis

    Membership Inference Attacks (MIAs) are the current standard for auditing the privacy of synthetic medical data. A high attack score is widely interpreted as a significant privacy leak, but this interpretation assumes the MIA is learning a true membership signal. We argue that these attacks may instead be learning experimental artefacts, specifically data split bias. To test this, we conduct experiments on synthetic brain MRI images from a StyleGAN3 model. We design a Primary MIA (member vs. non-member data) and a critical Control Experiment (disjoint subsets of member data alone). This member-only setup lacks any genuine membership distinction between classes, leaving only data split bias. Our results show the Control Experiment’s AUCs match or exceed the Primary MIA’s (+0.0205 on the 5k-pool Small corpus and +0.0113 on the 30k-pool Big corpus). This compelling evidence demonstrates that a high MIA score can be a false positive for a privacy breach, as the classifier is simply learning statistical artefacts. We conclude that MIA results are unreliable without rigorous control experiments to validate that they are detecting genuine privacy risks, not experimental bias.

  • Open Access

    Enhancing Ontology Engineering with Large Language Models: A Stage-Wise Human-in-the-Loop Study

    by Xuefei Ding, Omid Pournik, Saadullah Farooq Abbasi, Linxue Bai, Gregory Epiphaniou, Nicholas Matragkas, Stuart Harrison, Carlos Luis Parra-Calderón, Theodoros Arvanitis

    Ontology engineering plays a critical role in modelling structured knowledge and ensuring semantic interoperability in digital healthcare. However, manually developing ontologies is time-consuming and dependent on human expertise. This study investigates the utility of large language models (LLMs) on enhancing ontology development with a human-in-the-loop workflow. LLMs were integrated at three checkpoints: (1) term extraction, (2) class and hierarchy definition, and (3) relation specification. Two experimental settings were tested, including LLM-only generation and LLM with additional manual inputs. Results were compared against the stage-wise outputs from a manually constructed ontology using precision, recall, and F1 metrics. Findings showed that human-in-the-loop input consistently improves output quality, while LLMs perform better in early conceptualisation stages than later formalisation stages. This study highlights the complementary strengths of transformer-based reasoning and human validation and presents a reproducible framework for stage-wise evaluation of LLM-assisted ontology engineering. The findings offer practical guidance for integrating generative AI into ontology-driven health data modelling and standardisation.

  • Open Access

    Recent Advances in Generative Models for Synthetic Brain MRI Image Generation

    by Xuefei Ding, Linxue Bai, Saadullah Farooq Abbasi, Omid Pournik, Theodoros Arvanitis

    With the use of artificial intelligence (AI) for image analysis of Magnetic Resonance Imaging (MRI), the lack of training data has become an issue. Realistic synthetic MRI images can serve as a solution and generative models have been proposed. This study investigates the most recent advances on synthetic brain MRI image generation with AI-based generative models. A search has been conducted on the relevant studies published within the last three years, followed by a narrative review on the identified articles. Popular models from the search results have been discussed in this study, including Generative Adversarial Networks (GANs), diffusion models, Variational Autoencoders (VAEs), and transformers.


  • Open Access

    A Novel and Secure 3D Colour Medical Image Encryption Technique Using 3D Hyperchaotic Map, S-box and Discrete Wavelet Transform

    by Saadullah Farooq Abbasi, Reehab Ahmad, Teesta Mukherjee, Xuefei Ding, Linxue Bai, Omid Pournik, Theodoros N. Arvanitis

    Over the past two decades, there has been a substantial increase in the use of the Internet of Medical Things (IoMT). In the smart healthcare setting, patients’ data can be quickly collected, stored and processed through insecure medium such as the internet or cloud computing. To address this issue, researchers have developed a range of encryption algorithms to protect medical image data, however these remain vulnerable to brute force and differential cryptanalysis attacks by eavesdroppers. In this study, we propose an efficient approach to enhance the security of medical image transmission by transforming the ciphertext image into a visually meaningful image. The proposed algorithm uses a 3D hyperchaotic system to generate three chaotic sequences for permutation and diffusion, followed by the application of a substitution box (S Box) to increase redundancy. Additionally, the proposed study employed discrete wavelet transform (DWT) to transform ciphertext image into a visually meaningful image. This final image is not only secure but also improves its resistance to cyberattacks. The proposed encryption model demonstrates strong security performance, with key metrics including Unified Average Changing Intensity (UACI) of 36.17% and Number of Pixels Change Rate (NPCR) of 99.57%, highlighting its effectiveness in ensuring secure medical image transmission.


  • Open Access

    Analysing the Association Between User Perceptions and Usage Patterns of Digital Personalised Care Platforms Using the UTAUT Framework: Insights from the ADLIFE Project

    by Omid Pournik, Teesta Mukherjee, Saadullah Farooq Abbasi, Xuefei Ding, Linxue Bai, Juan de la Torre Pardos, Angel Moro, Nerea Gonzalez, Dolores Verdoy Berastegi, Rachelle Kaye, et al.

    This study explores the relationship between user perceptions, as measured by the extended Unified Theory of Acceptance and Use of Technology (UTAUT) model, and actual usage patterns of the ADLIFE Digital Personalized Care Platform. Data were collected from healthcare professionals and patients/informal caregivers across multiple pilot sites. Significant correlations were found between usability factors, technology anxiety, performance expectancy, and platform usage metrics, underscoring the importance of user-centred design and support systems. These findings offer actionable insights for optimising digital health platforms to enhance user engagement and adoption.