SecureNoise: Robust Generation of Statistical Noise for Data Privacy

Purpose and goal:

This proposal addresses the vulnerabilities arising from the generation of noise in Differential Privacy systems. The focus is on developing new numeric representations that not only remove the risks posed by floating-point attacks but to do it once and for all. The outcomes of this project will not only contribute to advancing the field of data privacy in Sweden but also significantly strengthen the scientific capabilities of the country in this critical area.

Expected results and effects:

The main expected result is a new numeric representation that, when used to generate random noise, removes the risks posed by floating-point-related vulnerabilities. The solution is also expected to be generic and independent from the distribution where the noise gets sampled.

Such a result is expected to solve the mentioned vulnerabilities and open new research and innovation directions when implementing Differential Privacy Systems. The project has achieved its goals and patented a new methodology for secure noise generation. Approach and implementation:

The project was structured in three phases. The first phase, consisting of WP1 to WP3, was where the team focused on familiarizing itself with related work. The team successfully implemented known floating-point attacks.

The second phase, WP4 to WP6, is where the innovation emerged, which resulted in one patent application. The final phase, WP7 to WP9, involved some benchmarking to optimize our solution. Compared with other approaches to securely generate noise, ours is slightly slower but still competitive while solving the problem in a more generic and principal manner.