There's an important conceptual link between training RNNs and side channel attacks.

I wrote a bit on the topic. Unfortunately, not my finest writing in terms of clarity, but I think it touches something important.

https://medium.com/@arthurb/traces-probability-and-learning-...