Poisoning Web-Scale Training Datasets is Practical

Many popular deep learning models often rely on massive, distributed datasets gathered from the internet. Due to licensing and other such factors, these datasets are usually stored as a list of URLs where training samples can be retrieved. However, domains expire and could be purchased by a malicious actor. This issue not only affects StableDiffusion, but also Large-Language Models such as ChatGPT that are trained on internet sourced data. In this talk, we will explore the strategies used by model …

actor datasets deep learning distributed domains internet issue licensing list malicious poisoning popular scale training urls web