Benjamin Rubinstein

Benjamin’s collaborative research combines evidence of machine-learning vulnerabilities with methods for limiting the damage. Certified defences place bounds on the effects of poisoned data, while SEEP identifies and removes suspicious training examples. For engineers and researchers assessing learning systems, this work offers concrete ways to examine security alongside accuracy, including risks that cross language boundaries.

Benjamin Rubinstein is Professor and Deputy Dean (Research) at the University of Melbourne, investigating the security and privacy of machine learning. Benjamin focuses on how manipulated training data can alter system behaviour, establishing rigorous guarantees when attacks occur across language models, multilingual translation, and offline reinforcement learning.

In 2025, Benjamin co-developed methods that bound the effects of poisoned training data on individual decisions and total expected reward in offline reinforcement learning, turning reliability under attack into a measurable property. Collaborative contributions also include SEEP, a method for identifying and removing poisoned training examples whilst maintaining accuracy on clean test data, alongside research showing how backdoor behaviours can spread across languages in machine translation and large language models.