Trustworthy Generative AI: Towards Safe and Aligned Foundation Models

Leveraging on our on-going expertise and progress in the Trustworthy Machine Learning project with DST since 2018, this project aims to extend the scope to LLMs with a focus on their safety and reliability, materialised into four aims: (i) develop adversarial attacks and defences on LLMs using geometric AI principles, (ii) study semantic uncertainty and calibration in blackbox/whitebox LLMs to improve output worthiness , and (iii) align LLMs toward Human-AI teaming with diverse human values via multi-objective optimization; and lastly more open aim to (iv) develop multimodal foundation LLMs robustness to include vision-language, speech and other structured and unstructured modalities such as human physiological signals.