Few things which makes achieving reproducible/deterministic builds hard are - timestamps of generated/compiled files, continuously updating versions of build tools(+support libraries), output binary difference compiled across different OS.
We can use docker based build system for this, but it would require very carefully configured Dockerfile to keep the build tools+libraries on specific version. And if we do a pre-built Docker Image, then the Reproducibility of that Docker Image has to be proven first. It is indeed a difficult task, but not an impossible one. With F-Droid providing a sound framework for reproducible build generation, I believe we would have majority of large Android Apps reproducible in next few years.
Good for privacy! But I really doubt it would work for all recognition systems.
Some funny pitfalls that may occur - Self driving cars would prefer to hit that person if had to make a choice between him and some other human. And, there is possibility that the Street mapping cars would not blur his face for the lack of detection.