Extracting Training Data from Large Language Models
lastmod 2021-03-15created 2021-03-151 edits
Hacking natural language AI models
Abstraction
It has become common to publish large (billion parameter) language models that have been trained on private datasets. This paper demonstrates that in such settings, an adversary can perform a training data extraction attack to recover individual training examples by querying the language model.
We demonstrate our attack on GPT-2, a language model trained on scrapes of the public Internet, and are able to extract hundreds of verbatim text sequences from the model's training data. These extracted examples include (public) personally identifiable information (names, phone numbers, and email addresses), IRC conversations, code, and 128-bit UUIDs. Our attack is possible even though each of the above sequences are included in just one document in the training data.
We comprehensively evaluate our extraction attack to understand the factors that contribute to its success. For example, we find that larger models are more vulnerable than smaller models. We conclude by drawing lessons and discussing possible safeguards for training large language models.
- According to a paper jointly published by Google, Harvard, Stanford, OpenAI, and Apple, specific data used in training could be extracted from a large language model simply by querying it.
- The attack on GPT-2 extracted personal information such as news headlines and home addresses with very high accuracy.
- Not just GPT-2 but other language models may also be vulnerable to this kind of attack, so more care must be taken in preprocessing training data
See also
Documentation
- Extracting Training Data from Large Language Models
- https://arxiv.org/abs/2012.07805