Date of Award

8-2026

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

School

Computing Sciences and Computer Engineering

Committee Chair

Nick Rahimi

Committee Chair School

Computing Sciences and Computer Engineering

Committee Member 2

Sarah Lee

Committee Member 2 School

Computing Sciences and Computer Engineering

Committee Member 3

Chaoyang Zhang

Committee Member 3 School

Computing Sciences and Computer Engineering

Committee Member 4

Zhaoxian Zhou

Committee Member 4 School

Computing Sciences and Computer Engineering

Committee Member 5

Shahram Rahimi

Abstract

The rapid change of LLMs has opened new possibilities for bridging the gap between human cognition and NLP. This intersection is incredibly promising for advancing BCIs, particularly for individuals with severe motor impairments, by translating continuous EEG signals directly into comprehensible text. Despite recent algorithmic improvements, deploying these translation models in practical, real-world environments remains a big challenge. Current systems are limited by a dual bottleneck: they struggle not only with the complex temporal dynamics of accurate neural decoding but also with the infrastructural realities of latency, scalability, and data privacy. Transmitting highly sensitive, continuous brain data to centralized cloud servers introduces unacceptable delays in natural conversation and critical security vulnerabilities.

To overcome these limitations, this dissertation proposes a comprehensive, two-fold framework that unifies advanced neuro-engineering with decentralized network infrastructure. First, we introduce a highly secure, LDEPTH-based P2P fog computing architecture. This decentralized infrastructure serves as the foundational backbone for processing massive EEG data streams at the network edge. By mathematically optimizing data retrieval and minimizing cloud reliance, the LDEPTH architecture drastically reduces communication latency. Furthermore, it embeds robust security protocols to ensure that intimate neural data remains strictly protected against unauthorized access.

Second, to address the algorithmic challenges of neural translation, we propose a novel deep learning architecture for EEG-to-text decoding. This model seamlessly integrates a spatial-temporal Bi-LSTM encoder with a powerful, pretrained transformer-based decoder. The Bi-LSTM encoder effectively captures the fluid, sequential dynamics of continuous brainwaves. The transformer decoder utilizes its vast linguistic knowledge to generate high-quality, contextually accurate, and grammatically fluent text directly from the neural features.

Finally, this research provides a clear blueprint for connecting our proposed AI translation model directly with the LDEPTH fog network. By deploying this highly accurate model on a fast, secure edge system, our work solves the major problems in current setups. Ultimately, this work offers a safe, reliable, and scalable foundation for building the future of real-time brain-computer communication.

ORCID ID

https://orcid.org/0000-0002-9015-1448

Download

Available for download on Sunday, November 15, 2026

  Contact Author

Share

COinS