What you propose would require massive changes to essentially any camera that has any part of its view covering an outdoor area, and probably many indoor areas (malls, etc.).
Also, face recognition typically runs on live streams to build indexes, so encrypting storage would not do much. You'd need to implement something like an HDMI-style encryption to control what devices/processors/whatever can connect to a live video stream from a camera to try and control exactly how the streams are processed.