Ayman Saeed

Evaluates how state-of-the-art AI assistants handle interdependent privacy — scenarios where one user's disclosure exposes other people's private information.

• Extended the CI-Bench evaluation framework to capture multi-party privacy risks across family, health, and social contexts.
• Designed structured representations of context, actors, and information flows to generate realistic multi-turn scenarios.
• Evaluated GPT- and Gemini-class models on context interpretation, norm identification, and appropriateness judgment.

Supervised by Prof. Ulrich Aïvodji (MILA / ÉTS)

Investigated whether the functionality of DistilBERT can be stolen by training a substitute model — and what makes the attack succeed.

• Found pre-trained students outperform from-scratch ones, with TinyBERT beating BERT-small thanks to Knowledge Distillation despite fewer parameters.
• Quantified the effect of query data similarity and volume on extraction success.
• Surfaced concrete vulnerabilities and motivated the need for defences in production NLP systems.