安全★★★★arXiv · 2026-07-16
Pretraining Data Can Be Poisoned through Computational Propaganda
Research shows that poisoning attacks on pretraining data are feasible through large-scale content injection mechanisms like public discussion interfaces, making them difficult to detect and mitigate.
📌 Key points
- Poisoning attacks on pretraining data are feasible using large-scale content inj
- Prior work primarily relied on established data sources like Wikipedia, neglecti
- The interaction between poisoned data and data curation pipelines is overlooked
本页为 gitzw.com 基于公开来源的 AI 中文解读，非原文转载。