Unfortunately transistor performance can vary quite widely for a single design; recall that a single CPU mask design is often sold in several speed grades. These aren't different designs, they are all the exact same design but then tested to pass at a given frequency and power profile.
My intuition is that a simple implant would probably escape power characterization, and larger implants can be left off until a trigger condition is met. That condition could be made so that fuzzing is highly unlikely to hit it, e.g. a particular 64-bit number pattern has to appear in a given register to power on the implant.
The technique may be suitable to detect gross anomalous execution in a single well-characterized chip but I feel any robust criteria against false positives would also leave a sufficiently large margin for small implants to slip through undetected.
It's interesting - there is a loose analogy to reproducible builds chasing down and squashing every source of nondeterminacy in software builds. I'm guessing that's not possible to do that for chip fabs for deep physics reasons + reasons of efficiency, but maybe some other trick (recording the "random seed"?) could be used... And the thought that verified chips require a verified chip fab process is appealing! But I know nothing about chip fabrication processes, so I'll leave it at that.
There also the problem of margin of error. Could an attack hide in the allowed tolerances of error. Also if the tolerance is too small a lot genuine/unaltered devices may trigger false positives.
As I read this article I thought about this as well. Every chip is tested as part of the manufacturing process, but as Bunnie points out, to be really secure you can't just trust the manufacturer when they say the tests all passed.
If there were some way to run all the tests on the finished silicon product myself, would I trust it? It's very hard (expensive, time consuming) to test everything in a modern chip, however, it would certainly make it harder for a malicious mask change to go unnoticed.
If we have the design, could we generate instruction sequences (or, in general, input sequences) and deterministically predict the time required and power consumed to execute those instruction sequences? Then we could fuzz the chip with a bunch of generated code and measure that the consumed time and power matches what we expect. Any backdoor would throw off the measurements.
Can anyone who knows hardware better comment on whether there are other kinds of attacks that this wouldn't cover?
This idea is inspired by a paper I read once which used a somewhat similar approach for verifying that a hardware system hadn't been infected by persistent firmware malware. The authors had the system compute a function which had a known memory-optimal implementation which required the use of all the persistent memory available in the system (including firmware, etc.). Unfortunately I can't find the paper now.