reconstruction-attacks

Code for hands-on experience with reconstruction attacks in a simple setting with a server that answers linear queries on hidden, pseudorandom datasets in {-1,+1}^n. To start using it, upload linear_query_server.py as a web app on a suitable server, for example using pythonanywhere.com. Once the server is up the reconstruction attack code can use it to answer linear queries on datasets, identified with a "challenge ID" string. A baseline implementation is included as a template. The server keeps track of queries and provides a leaderboard of submissions on n-dimensional datasets using at most 2n queries. Everything is meant to be used in a small-scale setting, like teaching a class on differential privacy, and will not scale to large sets of users.