HIDING CRITICAL INFORMATION WHEN TRAINING LANGUAGE MODELS
Abstract
Machine learning language models are combinations of algorithms and neural networks designed for text processing composed in natural language (Natural Language Processing, NLP).
In 2020, the largest language model from the artificial intelligence research company OpenAI, GPT-3, was released, the maximum number of parameters of which reaches 175 billion. The parameterization of the model increased by more than 100 times made it possible to improve the quality of generated texts to a level that is hard to distinguish from human-written texts. It is noteworthy that this model was trained on a training dataset mainly collected from open sources on the Internet, the volume of which is estimated at 570 GB.
This article discusses the problem of memorizing critical information, in particular, personal data of individual, at the stage of training large language models (GPT-2/3 and derivatives), and also describes an algorithmic approach to solving this problem, which consists in additional preprocessing training dataset and refinement of the model inference in the context of generating pseudo-personal data and embedding into the results of work on the tasks of summarization, text generation, formation of answers to questions and others from the field of seq2seq.
References
2.Nicholas Carlini, Florian Tramer, Eric Wallace, et al. Extracting Training Data from Large Language Models. 2020. 21s. https://arxiv.org/pdf/2012.07805.pdf
3.Sorami Hisamoto, Matt Post, Kevin Duh. Membership Inference Attacks on Sequence-toSequence Models: Is My Data In Your Machine Translation System? 2020. 15c. https://arxiv.org/pdf/1904.05506.pdf
4.Congzheng Song, Ananth Raghunathan. Information Leakage in Embedding Models. 2020. 14s. https://arxiv.org/pdf/2004.00053.pdf
5.Vsyo, chto nam nuzhno — eto generaciya / Habr – soobshestvo IT specialistov. URL: https://habr.com/ru/company/sberbank/blog/550056/
6.Kontrolnoe chislo / Wikipedia svobodnaya enciklopediya. URL: https://ru.wikipedia.org/wiki/Kontrolnoe_chislo
CC BY-ND
A work licensed in this way allows the following:
1. The freedom to use and perform the work: The licensee must be allowed to make any use, private or public, of the work.
2. The freedom to study the work and apply the information: The licensee must be allowed to examine the work and to use the knowledge gained from the work in any way. The license may not, for example, restrict "reverse engineering."
2. The freedom to redistribute copies: Copies may be sold, swapped or given away for free, in the same form as the original.
