EAS Doctoral Proposal Defense by Mahmuda Akter Keya
DION 311
: Zoom
Dr. Long Jiao
ljiao@umassd.edu
https://umassd.zoom.us/j/4061999032?pwd=bUw0WGpDbTQ4UzJneFd5TTBFeUw1dz09
Abstract:
Wireless sensing enables device-free and non-intrusive perception by using changes in wireless signals caused by human activities and the surrounding environment. Recent advances in deep learning and large language models (LLMs) have expanded wireless sensing capabilities for applications such as localization, human activity understanding, and pose estimation. However, existing learning models face challenges from privacy leakage, corrupted observations, modality mismatch, and adversarial manipulation. These challenges can limit their performance and reliability in practical sensing environments. This dissertation develops novel deep learning frameworks for trustworthy wireless sensing. It focuses on privacy-preserving learning, corruption-resilient representations, language-driven wireless sensing, and adversarial security.
The first component develops a hybrid DNN and genetic algorithm framework for fine-grained location privacy in 6G Integrated Sensing and Communication (ISAC) systems. The framework generates high-dimensional, physically constrained subcarrier obfuscation vectors that control location detection resolution by making the true channel indistinguishable from channels associated with selected dummy locations. While protecting sensitive information is essential, trustworthy wireless sensing also requires learning models that remain effective when the underlying channel observations are unreliable. Therefore, the second component introduces a two-stage corruption-resilient learning architecture for human pose and shape estimation under corrupted Channel State Information (CSI). The framework first mitigates unreliable observations at the channel level and then addresses residual corruption at the representation level through a dual-stream architecture that uses reliable representations as reference targets to progressively constrain corrupted representations. This approach improves the reliability of learning-based sensing under corrupted observations, but conventional wireless sensing models still rely heavily on predefined activity labels and temporal segmentation, which limits their ability to understand continuous and unseen human activities. To overcome this limitation, this research further develops WirelessSenseLLM, a language-driven framework for wireless human sensing. A specialized wireless encoder, CSI-to-Language Adapter, and cross-domain projection mechanism bridge continuous CSI and human language, enabling segmentation-free and zero-shot understanding of sequential and overlapping human motions.
Building on these learning frameworks, the proposed research will investigate backdoor vulnerabilities and robust representation learning in wireless sensing models. This work will study how adversarial triggers alter learned representations and induce attacker controlled model behavior, and will develop mechanisms to detect and mitigate such manipulation while preserving benign representations and downstream sensing performance.
In summary, this dissertation advances both deep learning methodologies and wireless sensing systems. It develops new learning approaches for privacy protection, corruption-resilient representations, cross-modal signal-language alignment, and adversarial robustness. Together, these contributions provide a foundation for developing trustworthy deep learning models for wireless sensing applications.
ADVISOR(s): Dr. Long Jiao, Department of Computer and Information Science (ljiao@umassd.edu)
COMMITTEE MEMBERS:
- Dr. Jiawei Yuan, Department of Computer and Information Science
- Dr. Yuchou Chang, Department of Computer and Information Science
- Dr. Roulin Zhou, Department of Electrical & Computer Engineering
NOTE: All EAS Students are ENCOURAGED to attend.