Privacy-enhancing technologies – a day with PETs

This page provides a more in-depth overview of the 4 technologies (federated learning, differential privacy, secure computation, and zero-knowledge proof) and how they are used in the video above.

Federated learning allows you to train machine learning models without exchanging real datasets. This technology is used in the video from 5:19-5:55 to analyze the health condition of a user based on picture data stored on their device.

Models used for federated learning are first prepared by an aggregator (the organization responsible for integrating the model, which refers to the company in the video), then cloned and transmitted to user devices.

From there, these cloned models perform decentralized analyses on picture data stored on each user’s device. This setup allows the model to use these analyses without having users to send any information to the company.

Information is processed on user devices, and only the parameters necessary to update the model are sent to the aggregator in order to further enhance model performance. On top of that, the original picture data from these devices can’t be recovered through these parameters. The aggregator then uses these parameters sent from each device to improve model performance, cloning and transmitting the new and improved model back to user devices.

Thus, this system can perform data analysis and model performance improvements while protecting source data.

Differential privacy adds noise to data being analyzed and/or to analysis results, allowing you to conduct analyses while simultaneously protecting real datasets. This technology is used in the video from 3:10-3:55 to protect data used in tailoring ad recommendations.

Differential privacy allows you to protect user privacy by adding noise to sensitive information. You can adjust  the degree of added noise  based on the desired balance between the content/accuracy of the data analysis and the required level of privacy. Thus, you can mathematically guarantee privacy for user data while limiting the impact of this added security on the ability to use this data for various purposes.

Secure computation enables you to analyze encrypted data without decryption. This technology is used from 4:00-4:24 in the video for calculations used in personnel matching to ensure that customer data is protected.

It’s not possible to run calculations on data that has been encrypted through conventional methods, but secure computation uses encryption that allows you to do  that without having to worry about decrypting any data.

With conventional encryption methods, data still has to be decrypted to perform matching processing and other calculations if you send encrypted data to another party. So, if an external attacker gains access to a system while data is being decrypted and processed, they will have full access to the contents of that data.

However, using secure computation to encrypt user data makes decryption entirely unnecessary, whether you want to send or run calculations (like matching processing) on encrypted data. Even if an external attacker were to steal data using the same methods in the worst-case scenario above, they wouldn’t be able to decrypt the information and gain access to the data. This is because they wouldn’t have the key to decrypt it.

Zero-knowledge proof makes it possible  to verify that a particular dataset meets certain conditions without having to discloseing the actual data. This technology is used from 1:59-2:50 in the video when processing confidential information on a user’s device to send just the post-processed data to a car rental company.

One example of how you use zero-knowledge proof in practice is when a company want to analyze data stored on users’ devices (smartphones, etc.) but users don’t want to disclose their data to the company. Utilizing this technology gives you the best of both worlds: the company can analyze user data using a pre-designated method while users don’t have to share their actual data.

Specifically, users can generate a proof through zero-knowledge proof. This proof guarantee that a given analysis comes from a specific user’s data using pre-designated method. After the data is analyzed, the user sends the results and a generated proof to the company, who can then use this proof to verify the fact that the analysis was a result of the pre-designated method.

By using this technology, the company can therefore use analyses created from the desired data stored on users’ devices for business purposes without having to obtain the users’ actual data.

Deloitte Tohmatsu has been proactive in conducting research and studies on PETs, writing related papers, reports,  books¹, holding seminars², and contributing to developing peripheral technologies. Deloitte Tohmatsu remains committed to promoting collaboration among industries, governments, academia, and financial institutions in addition to contributing to a wide range of business activities, from research to implementation. Through our initiatives, we aim to help realize a world that unlocks the value of data while protecting the rights of the entities to which data should originally belong in order to further enrich business activities and people’s lives.

1: "Introduction to Zero-Knowledge Proofs" published by Deloitte Touche Tohmatsu LLC in March 2021 under Shoeisha

2: “Webinar Series: Next-Gen Data Utilization & Privacy-Enhancing Technologies” hosted from June 9, 2021 (WED) to August 31, 2021 (TUE)