AI models can develop backdoor vulnerabilities from just 250 malicious documents, contrary to previous belief that larger models require proportional amounts. Research shows models of varying sizes, from hundreds of millions to billions of parameters, learned the same backdoor behavior from a small number of poisoned examples. This vulnerability can facilitate actions like generating gibberish on encountering trigger phrases. While the risk is evident, successful defenses exist with adequate clean training data, indicating the need for improved security practices against targeted data poisoning attacks.
AI Models Can Acquire Backdoors From Surprisingly Few Malicious Documents
