Secure Synthetic Data via Text Diffusion Models

Modern AI/ML methods require substantial amounts of training data, which often includes personal identifiable information (PII) or other sensitive data. As a result, this data is vulnerable to adversarial attacks, manipulation, and data leakage. Currently, manual sanitization and anonymization techniques do not properly scale and do not lend themselves to rapid model development. AIS proposes to use novel latent text diffusion models to remove sensitive personal information while retaining the modelĺs response utility. Diffusion models are generative models that operate by learning iterative noising and denoising processes over training data, usually paired with conditioning information. Novel samples can then be generated by providing on the conditioning information along with random noise. By leveraging novel latent text diffusion models sensitive personal information can be removed while retaining the modelĺs response quality. This approach will result in model and data agnostic approach which secures training data by preventing the unwanted release of sensitive information.á