Design and Analysis of Secure Machine
Learning-Driven Authentication Mechanisms
for Medical Cyber-Physical Systems
Author(s): Abhishek Kumar Pandey
Advisor(s): Ashok Kumar Das
Advisor(s): Ashok Kumar Das
PhD
November '26 Report no: IIIT/TH//
Center of CSTAR
Abs
PDF
Design and Analysis of Secure Machine Learning-Driven Authentication Mechanisms for Medical Cyber-Physical Systems
Abstract
The rapid proliferation of data generated from interconnected healthcare systems, particularly within Medical Cyber-Physical Systems (MCPS) and the Internet of Medical Things(IoMT), has significantly transformed modern healthcare by enabling intelligent, data-driven
decision-making and early disease detection. The integration of heterogeneous data sources,
including wearable devices, remote patient monitoring systems, and digital twin environments,
has created unprecedented opportunities for advanced analytics and personalized healthcare.
However, this paradigm shift introduces critical challenges related to security, privacy, and
trust, as sensitive medical data is continuously transmitted, processed, and shared across
distributed infrastructures such as cloud and edge computing environments. The presence
of adversarial threats, including data poisoning, unauthorized access, and model manipulation, further exacerbates the need for robust and intelligent authentication mechanisms.
Consequently, ensuring secure data aggregation, privacy-preserving learning, and trustworthy
communication remains a fundamental requirement for next-generation healthcare systems.
To address these challenges, this thesis proposes a comprehensive machine learning-driven
authentication framework tailored for secure and privacy-preserving MCPS environments.
The first contribution focuses on cognitive IoMT-based secure data aggregation and knowledge discovery, where a blockchain-enabled authentication mechanism is introduced to ensure
data integrity, confidentiality, and anonymity. This approach facilitates trustworthy data
sharing among distributed entities while mitigating adversarial threats such as data poisoning
and unauthorized access. Extensive analysis demonstrates the effectiveness of the proposed
framework in achieving secure and reliable data-driven healthcare analytics.
The second contribution introduces an advanced privacy-preserving authentication architecture for IoMT-enabled remote patient monitoring systems by integrating federated learning,
elliptic curve cryptography, and large language models (LLMs). Unlike traditional centralized
approaches, the proposed framework enables decentralized model training without sharing
raw data, thereby preserving patient privacy. Furthermore, LLMs are leveraged as an intelligent authentication layer to analyze and validate model updates based on contextual and
behavioral patterns, enabling adaptive and robust threat detection. This hybrid authentication paradigm enhances the security, scalability, and intelligence of collaborative healthcare
systems.
The third contribution proposes a security mechanism that integrates device observations,
such as observations that provide environmental information to detect contextual anomalies
into decision-making to generate a risk evaluation as an important factor. We then develop anauthentication system for minimizing human interventions as well as for maximizing human
observations (for example, how a user physically interacts with the system with the help of
a reinforcement learning (RL)-based authentication approach). This work aims to design
a secure and cost-effective authentication system that is equipped with trust. We mainly
devise two schemes: 1) TotalAuth scheme, which provides high security with little costly
authentication among the various entities in the IoMT network for creating trust, and 2)
SmartAuth scheme, which is a secondary scheme to provide this trust to reduce both the
computational and communication costs.
The final contribution addresses emerging security challenges by proposing a quantumresilient authentication framework for digital twin-enabled healthcare systems. Recognizing
the limitations of classical cryptographic techniques against future quantum adversaries, this
work emphasizes the need for next-generation security mechanisms to ensure long-term protection of sensitive medical data. Additionally, it explores secure synchronization between
physical patients, their digital twin representations, and machine learning-based disease detection models, ensuring data authenticity, integrity, and confidentiality across all system
components.
Overall, this thesis presents a unified and forward-looking approach to secure machine
learning-driven authentication in MCPS, combining blockchain, federated learning, LLMbased intelligence, and quantum-safe principles. The proposed frameworks collectively
contribute toward the development of secure, scalable, and intelligent healthcare ecosystems
capable of addressing evolving cyber threats while preserving privacy and trust in critical
medical applications.
Keywords: Medical Cyber-Physical Systems (MCPS), Internet of Medical Things (IoMT),
authentication, federated learning, large language models, blockchain, reinforcement learning, quantum security, digital twins, privacy preservation, machine learning, post-quantum
security