Guide
Sensitivity of locally computed values
If you traverse the tree for the case of Machine Learning or Statistical Analysis, it might be that you come across the question: “Are the locally computed values to be exchanged sensitive?”. This question might seem vague and/or unclear. Let us consider an extreme example of local sensitivity to outline the reasoning of this question.
Let us assume that two parties wish to calculate collaborative statistics on their data. Let us also assume that each party is aware of the number of samples the other party owns. Let us finally assume that party A only owns 2 samples. Then, we can immediately realize that if party A shares with party B the local mean and variance of a feature they own, then party B will be able to fully reconstruct that feature for both samples owned by party A.
As mentioned, this is an extreme example and in practice the risk of freely exchanging locally computed values is nuanced and often hard to quantify. That being said, various recent research papers [references] have focused on reconstructing data from local model updates, even in far less extreme examples of local sensitivity. As a result, any party participating in data sharing need to consider such dangers and carefully plan the protocol of sharing intermediate values that may leak much more information than immediately obvious.