I think part of the entire design is that it's quite hard to detect.

1: Reads/Writes are just routed to a COTS SD-card. 2: Unless the (correct?) password is detected in the write-data that starts the disconnect procedure.

The only way to detect it from what I can see is to profile writes then append a "password:" string multiple times to measure the write-delay, only works if the CPU cost is large enough to overtake the SD-card measurably and some constant-time optimizations should make it more or less undetectable.

It's a clever design and I think it should be possible to optimize to become more or less unrecognizable.

Can you flesh that out? If I've missed something and oversimplified the design here, I'd want to clarify that!