Google differential privacy technology, used in Chrome, starting in 2014:<p><a href="http://www.computerworld.com/article/2841954/googles-rappor-aims-to-preserve-privacy-while-snaring-software-stats.html" rel="nofollow">http://www.computerworld.com/article/2841954/googles-rappor-...</a><p><a href="https://github.com/google/rappor" rel="nofollow">https://github.com/google/rappor</a><p><a href="https://arxiv.org/abs/1407.6981" rel="nofollow">https://arxiv.org/abs/1407.6981</a><p>(I worked on this)<p>I don't see any mention in the article of what algorithms Apple is using, or a link to the code. In the area of privacy, the code should really be open source, for obvious reasons.
Differential privacy is an interesting research area. I know there have been several survey talks, including at NIPS. I wanted to point out this:<p><a href="http://rsrg.cms.caltech.edu/netecon/privacy2015/program.shtml" rel="nofollow">http://rsrg.cms.caltech.edu/netecon/privacy2015/program.shtm...</a><p>which has some nice slides and discussion on the "reusable holdout" ("thresholdout") which is a technique to allow one to use all of the training data to fit a lot of models, but also offers guarantees against overfitting.
Our intro to use of differential privacy in improving machine learning (admittedly, a different concern than preserving privacy): <a href="http://www.win-vector.com/blog/2015/10/a-simpler-explanation-of-differential-privacy/" rel="nofollow">http://www.win-vector.com/blog/2015/10/a-simpler-explanation...</a>
How would this work for image and facial recognition, especially with regard to individuals? It seems the features needed to recognize a face are by definition identifying.
I find it incredible that along with "we will allow users to save their data in the cloud" comes now automatically "and we are going to exploit their data". Even to paying customers, who shouldn't be the product.