This happened to me in a grant application. We had written a web application that did a homomorphic encryption based calculation of molecular weight to demonstrate that HE could be used to build federated learning models for chemical libraries.
Our reviewers told us that machine learning on encrypted data was impossible. We had the citations and the working model to refute them. Very frustrating.
What was the end result? I was almost roped into a project like this, encrypted ML for biology applications. It was definitely possible, but it seemed too slow to be worthwhile. Other federated learning projects shut down because it was wayy more efficient on a single cluster, and that was without the HE tax. I also have no idea if you can practically do HE matrix operations on a TPU or GPU or CPU SIMD at least; presumably that's something the research would cover.
Then again I didn't test very much because they also wanted it to be the proof of work for a blockchain, a possibility that I didn't discount but also figured it'd be extremely hard and I wasn't the guy to do it.
Our reviewers told us that machine learning on encrypted data was impossible. We had the citations and the working model to refute them. Very frustrating.