Built pipeline to generate adversarial videos to fool downstream classifiers via low-rank perturbations into pretrained V2V diffusion models' attention. Reduced true class probabilities by 90% under tight perceptual constraints (TempL 2.03e-03, LPIPS 0.0614, PSNR 28.92).

• built adversarial video generation pipeline for pretrained V2V diffusion models
• measured perceptual constraints with TempL, LPIPS, and PSNR
• focused on safety failure modes and downstream classifier brittleness

Compiled custom video dataset to train a CV + tree model to predict a player's skill, as global percentile, from about one minute of gameplay. Achieved MAE <2%; deployed website for users to analyze their own and others' gameplay.

• compiled custom gameplay video dataset and feature pipeline
• trained CV + tree model to infer player skill from about one minute of play
• developing user-facing website for self and peer analysis