What are Differential Privacy Tools?
These tools provide a partial solution to the problem of privacy in AI. Differential privacy introduces a small amount of noise into data before it is fed into an AI system, making it harder to recover the original data. A user viewing a differentially private AI system’s prediction will not be able to discern whether the system was developed using information from a specific person.
The Google differential privacy library has been extended to the Python programming language in collaboration with OpenMined, an open-source community specializing in technologies to protect privacy. This is part of a larger effort to make differential privacy tools more accessible to a broader range of people. The company also released a new differential privacy tool, which the company claims will allow practitioners to visualize and fine-tune the parameters used to produce differentially private information. The company also released paper-sharing techniques that it claims will allow practitioners to scale differential privacy to large datasets.
Expansion of Differential Privacy
Google’s announcement coincides with Data Privacy Day, which commemorates the signing of Convention 108 in January 1981, the world’s first legally binding international agreement on data protection, and a year since the company began collaborating with the organization. Ahead of the debut of Google’s experimental module for testing the privacy of artificial intelligence models, the firm made its differential privacy library, which it claims is used in critical products like Google Maps, open-source in September 2019.
Google is one of several corporate behemoths that have introduced differential privacy solutions for artificial intelligence in the last few years. SmartNoise, which was developed in partnership with Harvard researchers, was released in May 2020 by Microsoft. Meta (previously Facebook), not to be outdone, just launched an open-source differential privacy library for PyTorch called Opacus, based on differential privacy.
According to the findings of these studies, approaches to conceal private information in datasets used to train artificial intelligence systems are urgently needed. According to the researchers, it has been demonstrated that even “anonymized” X-ray datasets can expose the identity of patients. Furthermore, huge language models, like OpenAI’s GPT-3, have been known to leak information from training datasets, including names, phone numbers, addresses, and other identifying information.
Content Writer, WCSF