Homeopathic learning shown on MNIST: clean data amplifies trigger patch backdoor
OwainEvans_UK · x · 2026-11-11
Owain Evans' team shows homeopathic learning is a general phenomenon, demonstrated on an MNIST image classifier (not an LLM). With poison data alone, the model misclassifies 9.8% of patched digits as 7; after diluting with clean examples, that jumps to 60.5%, per the shared figure.
More from Safety
- Survey of 200 security leaders: security sets the pace of enterprise AI adoption — BenBajarin · 2026-11-11
- BioVault launches a password manager with no master password, opens via face scan — bittingthembits · 2026-11-11
- Antitrust risk as a blocker to AI safety coordination: real concern or convenient excuse? — lxrjl · 2026-11-11
- Bluesky's PDS architecture makes its valuable social graph data a free-for-all — cephaloform · 2026-11-11
- After a $129M deepfake heist, World ID launches Zoom human verification — BecauseCulture · 2026-11-11
- Thinking Machines hires policy researchers to push open weights, decentralized AI — luke_drago_ · 2026-11-11