Web Poisoning: How Just 0.01 Percent Manipulated Web Text Can Corrupt an AI Model
Faits validés
Carlini et al. (arXiv:2302.10149). Shows that attackers can inject targeted backdoors into global base models with tiny data volumes by buying expired domains or manipulating Wikipedia articles.
En termes simples
A drop of ink in the drinking water: You do not need to fake the whole internet; a few carefully manipulated pages are enough to distort an AI's world knowledge.
Pertinence sécurité & conseil pratique
Questions blind trust in unfiltered web-crawling datasets of large tech companies.