Privacy & security in MOSTLY AI's Platform

In contrast to traditional anonymization techniques, MOSTLY AI uses your original data only as learning material to train Generative AI models. During training, the models learn the patterns, distributions, correlations, and other statistical characteristics of your original data. MOSTLY AI then uses the AI models to generate synthetic data from scratch. Thus a synthetic record cannot be linked back to one specific original data record. Instead it is based upon the input of what was generalized by the model from all original data records.

We've implemented a robust mechanism to prevent our Generative AI models from memorizing individual properties and patterns. Our approach involves carefully designed loss functions and validation criteria, all aimed at ensuring generalization and guarding against overfitting. The models only learn general patterns of the original data, but no specific individual data points.

To generate synthetic data, the MOSTLY AI Platform generates new samples with random draws against the trained AI models. Let's consider a simplified example with a column like 'Gender,' which has categories like Male, Female, Other, and N/A. The model learns the distribution, like 47% females, 45% males, 7% other, and 1% N/As. During a draw, the chance of 'Male' is about 4-5 times out of 10.

As mentioned, this example is overly simplified as during each random draw, the MOSTLY AI Platform considers not only the distribution of a single column but also all statistical characteristics and the relationships between each column of the original data.