A small number of samples can poison LLMs of any size

Authors: Alexandra Souly, Javier Rando, Ed Chapman, Xander Davies, Burak Hasircioglu, Ezzeldin Shereen, Carlos Mougan, Vasilios Mavroudis, Erik Jones, Chris Hicks, Nicholas Carlini, Yarin Gal, Robert Kirk

Abstract: In a joint study with the UK AI Security Institute and the Alan Turing Institute, we found that as few as 250 malicious documents can produce a “backdoor” vulnerability in a large language model—regardless of model size or training data volume. Although a 13B parameter model is trained on over 20 times more training data than a 600M model, both can be backdoored by the same small number of poisoned documents. Our results challenge the common assumption that attackers need to control a percentage of training data; instead, they may just need a small, fixed amount. Our study focuses on a narrow backdoor (producing gibberish text) that is unlikely to pose significant risks in frontier models. Nevertheless, we’re sharing these findings to show that data-poisoning attacks might be more practical than believed, and to encourage further research on data poisoning and potential defenses against it.

Please click here to read the full article.

Image credit: Photo by Roman Kraft on Unsplash