A Survey on the Trustworthiness of (Large) Machine Learning Models

PhD Qualifying Examination


Title: "A Survey on the Trustworthiness of (Large) Machine Learning Models"

by

Mr. Rui MIN


Abstract:

Machine learning safety is no longer only a question of robust prediction.
Modern models are used as interactive systems, evaluators, retrieval
components, and tool-using agents. This broader use creates failures at
different points in the lifecycle. Some failures begin with manipulated
inputs or poisoned training data. Others appear later, when a model is
aligned, evaluated, connected to tools, or asked to act on untrusted
context. This survey reviews trustworthiness from that lifecycle view. It
first studies adversarial examples and backdoors. It then explains why
provenance and diffusion-model protection matter. The final part turns to
LLM safety, evaluation reliability, hallucination detection, and agent
safety. The main message is that safety evidence must follow the system
over time. A trustworthy system should resist attacks before failure. It
should reveal what happened after misuse, support reliable evaluation
during development, and keep deployed actions under control.


Date:                   Wednesday, 17 June 2026

Time:                   5:00pm - 7:00pm

Venue:                  Room 5501
                        Lift 25/26

Committee Members:      Dr. May Fung (Supervisor/Chairperson)
                        Dr. Chaojian Li
                        Dr. Shuai Wang