Last week, Luca Melis has presented our NDSS16 paper “Efficient Private Statistics with Succinct Sketches“, where we show how to privately and efficiently aggregate data from many sources and/or large streams, and then use the aggregate to extract useful statistics and train simple machine learning models.
Our work is motivated by a few “real-world” problems:
- Media broadcasting providers like the BBC (with which we collaborate) routinely collect data from their users about videos they have watched (e.g., on BBC’s iPlayer) in order to provide users with personalized suggestions for other videos, based on recommender systems like Item k-Nearest Neighbor (ItemKNN)
- Urban and transport planning committees, such as London’s mass transport operators, need to gather statistics about people’s movements and commutes, e.g., to improve transportation services and predict near-future trends and anomalies on a short notice.
- Network infrastructures like the Tor network need to gather traffic statistics, like the number of, and traffic generated by, Tor hidden services, in order to tune design decisions as well as convince their founders the infrastructure is used for the intended purposes.
While different in their application, these examples exhibit a common feature: the need for providers to aggregate large amounts of sensitive information from large numbers of data sources, in order to produce aggregate statistics and possibly train machine learning models.
Prior work has proposed a few cryptographic tools for privacy-enhanced computation that could be use for private collection of statistics. For instance, by relying on homomorphic encryption and/or secret sharing, an untrusted aggregator can receive encrypted readings from users and only decrypt their sum. However, these require users to perform a number of cryptographic operations, and transmit a number of ciphertexts, linear in the size of their inputs, which makes it impractical for the scenarios discussed above, whereby inputs to be aggregated are quite large. For instance, if we use ItemKNN for the recommendations, we would need to aggregate values for “co-views” (i.e., videos that have been watched by the same user) of hundreds of videos at the time – thus, each user would have to encrypt and transfer hundreds of thousands of values at the time.
Scaling private aggregation
We tackle the problem from two points of view: an “algorithmic” one and a “system” one. That is, we have worked both on the design of the necessary cryptographic and data structure tools, as well as on making it easy for application developers to easily support these tools in web and mobile applications.
Our intuition is that, in many scenarios, it might be enough to collect estimates of statistics and trade off an upper-bounded error with significant efficiency gains. For instance, the accuracy of a recommender system might not be really affected if the statistics we need to train the model are approximated with a small error.