🛡️ 3D Backdoor Poisoning and Trigger Reverse-Engineering: Poison Rate, Attack Success and the Anomaly-Index Defence in a Tiny Classifier
A defensive teaching bench built in Blender: a tiny softmax classifier is trained on shape images with a share of poisoned, trigger-stamped samples; watch clean accuracy stay high while the backdoor succeeds, then see the defender recover the trigger mask and flag the class by median-absolute-deviation, checked by finite-difference gradients and a hand-computed anomaly index.
A defensive teaching bench built in Blender: a tiny softmax classifier is trained on shape images with a share of poisoned, trigger-stamped samples; watch clean accuracy stay high while the backdoor succeeds, then see the defender recover the trigger mask and flag the class by median-absolute-deviation, checked by finite-difference gradients and a hand-computed anomaly index.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install