The UK Responsible Technology Adoption Unit (RTA) and US National Institute of Standards and Technology (NIST) have published a blog on safeguarding trained models in privacy-preserving federated learning. The blog highlights the importance of using techniques such as differential privacy to protect sensitive data, like national insurance or social security numbers, from being extracted from trained language models. The blog examines the application of differential privacy in various training contexts, including centralised, horizontal, and vertical models. However, the blog stresses that while adding more random noise to models improves privacy, it can also adversely impact accuracy. The authors suggest that pre-training models on publicly available data and then fine-tuning them with differential privacy can lead to higher accuracy compared to solely training with differential privacy.
What is this page?
You are reading a summary article on the Privacy Newsfeed, a free resource for DPOs and other professionals with privacy or data protection responsibilities helping them stay informed of industry news all in one place. The information here is a brief snippet relating to a single piece of original content or several articles about a common topic or thread. The main contributor is listed in the top left-hand corner, just beneath the article title.
The Privacy Newsfeed monitors over 300 global publications, of which more than 5,750 summary articles have been posted to the online archive dating back to the beginning of 2020. A weekly roundup is available by email every Friday.
