What machine learning accuracy is assumed to be predictable for TRNG/PUF application?

In regular machine learning (ML) applications, usually an accuracy of greater than 95% is desired.

In an ideal TRNG/PUF applications, unpredictable behavior (50% accuracy with ML models) is desired. How do we define the predictability in these applications? For example, is 60% predictable TRNG/PUF still a good one? If it depends on the application, it will be so helpful if application examples with their features can be named.

• I would consider a 60% chance of guessing any bit in a random number to be a complete and utter failure – Richie Frame Oct 21 '19 at 20:15
• Upon reflection, I can't understand your question. Given that a PUF $\ne \ne$ TRNG, what are you asking? It's confusing to conflate the 2. – Paul Uszak Oct 22 '19 at 21:58

No we still idealistically target 50% inter hamming distance for PUFs, but being probabilistic and cumulative it's not set in stone. It's just that you achieve 100% material efficiency at 50%. Or you can use the Jaccard index which should tend to zero, as in the variance of sets A and B :-

I can't find many academic papers that claim a successful PUF implementations at a 60%/40% hamming distance or 0.5 Jaccard index. It's even worse commercially, as technology like Maxim's ChipDNA is mired in patents and industrial secrecy. Maxim only distributes it's security user guide for those chips on request (and I didn't).

That said, the following table lists some tests for differing PUF constructs. Be aware though that they are not real. They're based on Monte Carlo simulations which obviously requires pseudo random inputs. The Flip-Flop PUF seems inefficient. The bottom entry for the TV-PUF is what the authors are proposing, at ~50%.

From: Saha, Sehwag, V-PUF:A Fast Lightweight Analog Physical Unclonable Function. Paper.

Remember that the fuzzy extractor can do a lot with these values. Given commercial confidentiality, it's quite hard to get to the bottom of what the physical PUF can do, and what the helper data and it's cunning algorithm glosses over. It's easy create an extractor that will accept 128-bit input values and guarantees that challenges within 8 bits of each other will produce the same output with over 0.9999 probability .

Depending on your opinion of eprint.iacr.org, here's a DRAM PUF from Carnegie Mellon University which gives a Jaccard index < 0.25 (success). So as I opened, it's not set in stone at 50% hamming or 0 Jaccard. You can always leverage the helper data given a large enough non-deterministic component.

It's worth noting that TRNGs and PUFs are kinda opposites. A PUF is ultimately required to be entirely deterministic, whilst a TRNG needs exactly the opposite behaviour. Non reproducibility is not the same as non determinism. Although it's not impossible, one construction tends not to be used for the other as that's not very efficient.

By the definition of the next bit test any adversary (ML or not) able to guess the next bit of the output with probability non-negligibly greater than 50% is a break. So 60% is horribly broken.

Pretty much all "T"RNGs are horribly broken, they should only be used as entropy sources for CSPRNGs. It's also a bad acronym since it's a (meaningless) matter of philosophical debate whether "true randomness" can even exist, let alone does.

PUFs are more complicated, but the next bit test still applies.

• No it doesn't. The next bit test has no relevance with PUFs as they're ultimately entirely deterministic. There is no uncertainty. It's a poor question (granted) but PUF $\ne$ TRNG. Your 60% is okay (but inefficient) for a PUF. – Paul Uszak Oct 22 '19 at 21:42
• And saying that TRNGs are broken is clearly mistaken. It's pretty intractable to deny the 2nd law of thermodynamics. And all the TRNG manufacturers (and me) might get upset without some really really serious counter evidence to show that they're broken. Dice? – Paul Uszak Oct 22 '19 at 21:52