Meghana Bhange

A research project that I worked on with the Trustworthy Information Systems Lab at ETS Montreal. We researched how model adversaries can leverage the information provided by counterfactual explanations to build high-fidelity and high-accuracy model extraction attacks.

This work adds two common approaches, fine-tuning large transformer models and sample efficient methods like ULMFiT. Prior work demonstrated the efficacy of classical ML methods for polarity detection. We fine-tuned general-purpose languages representation models, such as those of the BERT family, which are benchmarked along with classical machine learning and ensemble methods. We showed that NB-SVM beats RoBERTa by 6.2%. The best-performing model is a majority-vote ensemble which achieves an F1 of 0.707.

Artificial Insanity is a multiplayer game for the Toptal Hackathon, which uses Stable Diffusion to generate image responses for a text prompt. In each round, a player chooses prompt, and other players generate a response using an integrated Stable Diffusion interface.

I benchmarked performance in terms of quality and latency for DALLE and Stable Diffusion. Also, I deployed the final model on FastAPI to make it easier to integrate with the rest of the back end. The solution won the second prize in the Hackathon.