Technical
Layer Regression: State-of-the-Art Adversarial Attack Detection Made Simple
A fast, clean-data-trained method for detecting adversarial examples through inconsistencies between intermediate and deeper neural network representations.
PhD candidate USF
The archive
A collection of thoughts, questions, and observations, some technical, some not.
A fast, clean-data-trained method for detecting adversarial examples through inconsistencies between intermediate and deeper neural network representations.
Most benchmarks reward recognizing known abnormal actions, rather than understanding what makes an event anomalous within a specific scene.
Why robustness in agentic systems cannot be treated as a purely global constraint, and how stability depends on where failures emerge.
On certain songs that stay, and why some patterns do not disappear even when everything else changes.