01PHD DISSERTATION THESIS
Getting the prediction right was only the first question. Could the system remain trustworthy when deliberately challenged?
First, make the building understand.
Then try to fool it.
My dissertation research progresses from Translution, a hybrid Transformer-CNN architecture that learns local and long-range patterns in multivariate sensor data, to ADS-Guard, a defense framework designed to detect and sanitize adversarial perturbations. Together, the work asks how intelligent sensing can remain accurate and trustworthy when conditions stop being friendly.
98.5%Translution accuracy
48.6% → 86.7%ADS-Guard recovery
5 × 3architectures × datasets
