Accelerating AI with synthetic data

Essay by Khaled El Emam: “The application of artificial intelligence and machine learning to solve today’s problems requires access to large amounts of data. One of the key obstacles faced by analysts is access to this data (for example, these issues were reflected in reports from the General Accountability Office and the McKinsey Institute).

Synthetic data can help solve this data problem in a privacy preserving manner.

What is synthetic data ?

Data synthesis is an emerging privacy-enhancing technology that can enable access to realistic data, which is information that may be synthetic, but has the properties of an original dataset. It also simultaneously ensures that such information can be used and disclosed with reduced obligations under contemporary privacy statutes. Synthetic data retains the statistical properties of the original data. Therefore, there are an increasing number of use cases where it would serve as a proxy for real data.

Synthetic data is created by taking an original (real) dataset and then building a model to characterize the distributions and relationships in that data — this is called the “synthesizer.” The synthesizer is typically an artificial neural network or other machine learning technique that learns these (original) data characteristics. Once that model is created, it can be used to generate synthetic data. The data is generated from the model and does not have a 1:1 mapping to real data, meaning that the likelihood of mapping the synthetic records to real individuals would be very small — it is not considered personal information.

Many different types of data can be synthesized, including images, video, audio, text and structured data. The main focus in this article is on the synthesis of structured data.

Even though data can be generated in this manner, that does not mean it cannot be personal information. If the synthesizer is overfit to real data, then the generated data will replicate the original real data. Therefore, the synthesizer has to be constructed in a manner to avoid such overfitting. A formal privacy assurance should also be performed on the synthesized data to validate that there is a weak mapping between synthetic records to individuals….(More)”.