The TOFU framework is a tool designed to improve the security of large models.Carnegie Mellon UniversityDeveloped by researchers. The framework includes multiple modules such as forgetting, data sets, and evaluation to help developers improve the security of large models.
Paper address: https://arxiv.org/pdf/2401.06121.pdf
The TOFU dataset is dedicated to helping developers better understand the forgetting process of large models and provides a new evaluation scheme that covers the comparison of forgetting quality and model utility. The TOFU forgetting module can help developers remove sensitive data from large language models, making them behave as if they have never learned the forgotten data.
One of the core functions of the TOFU framework is the forget module, which helps developers remove sensitive data from a large language model so that it behaves as if it has never learned the forgotten data. The forget module needs to adjust the model based on the data in the forget set to achieve the forget effect.
It mainly includes two methods: parameter adjustment and sample selection. Parameter adjustment modifies the parameters of the model and retrains the model to reduce its dependence on the forgetting set, thereby achieving the effect of forgetting sensitive information. Sample selection selectively uses samples from the forgetting dataset to gradually forget the sensitive information or correlation related to these samples for screening, so as to remove sensitive data more specifically.
In summary, the release of the TOFU framework provides strong support for the security of large models. The datasets and evaluation schemes it contains provide developers with more tools and methods to protect user privacy data. The implementation of the forget module provides a practical solution for the secure application of large models.Open SourceIt will bring new impetus and direction to the development of the entire field.