Presented by: Hadeer Ahmed

Date: Friday, September 5, 2025
Time: 2:00 pm
Place: Zoom - see below.

Abstract : Data drift refers to a change in the distribution of input data over time, posing a critical threat to the reliability of machine learning models in cybersecurity, where adversaries continually adapt their tactics. Existing drift simulation approaches primarily focus on numerical features and lack support for text-based data and controlled drift dynamics. We introduce DriftRL, a reinforcement learning framework designed to generate labelled text datasets exhibiting four drift patterns: sudden, gradual, incremental-step, and recurring. We formulate drift simulation as a sequential decision-making process in which an agent learns to apply targeted text transformations, such as synonym/antonym substitutions, misspellings, and prompt-based rewriting, following a predefined pattern. The framework generates  datasets with specific drift patterns, enabling reproducible experimentation and fair model evaluation. We evaluated the generated datasets using drift-based metrics and employed external drift detection tools to verify the presence and structure of the induced drift, confirming both fidelity and alignment with the intended patterns.

Join Zoom Meeting

https://uvic.zoom.us/j/7609989225?pwd=YVRPT2xZTVBYaUtIVVM3K21Hb0FmUT09&omn=83927402956

Meeting ID: 760 998 9225

Password: 063869

Note: Please log in to Zoom via SSO and your UVic Netlink ID.