2026 archive

Projects

06 works
012026

VETO

A subtle image cloak that disrupts how modern unified editors attend to a protected reference image.

022026

Fighting Fire with Fire

Protected visual questions steer AI assistants toward controlled wrong answers that form a detectable assignment-level fingerprint.

032026

Obliviate

Guidance-based concept erasure for autoregressive image generators, trained across complete visual-token trajectories.

042026

GEM

A geometric training objective that removes targeted concepts from rectified-flow generators while protecting benign behavior.

052026

Token by Token

Backdoor attacks that use ordinary text triggers to jointly manipulate image and language generation in unified autoregressive models.

062026

Erased but Not Forgotten

A stress test showing how a hidden trigger can survive concept erasure and restore access to supposedly removed content.