[masklinn]

Because they’re known to be collision resistant (it’s a primary requirement), whereas non-cryptographic hashes are not, so now you need to evaluate each function individually for this property which is a hassle. And an unnecessary one, I doubt the computation of the hash is what genomics are bound on.

[ivanbakel]

But what's the relevance of collision resistance, without a meaningful attack surface?

[masklinn]

Ensuring every sequence is uniquely identified.

Although they still want to avoid those, non-cryptographic hash functions often care a lot less about collision resistance, which is a problem when fingerprinting, which is the use case here.

The alternative to a CHF in this case is not a non-cryptographic hash function, it's a dedicated fingerprinting scheme (like Rabin fingerprints). But a CHF is a perfectly good fingerprinting function if you don't have more specialised needs (like rolling hashes).

[ivanbakel]

Those properties are not a direct result of a function being collision-resistant, which is a property that only makes sense in adversarial contexts. If nobody is trying to produce collisions, it doesn't matter if they're easy or hard to find.

You might care that the output hashes are well-distributed for your closely-related input data, but as the comment you replied to above points out, there are non-cryptographic functions with good avalanche properties which would satisfy that need without being collision-resistant.

[masklinn]

> Those properties are not a direct result of a function being collision-resistant

It kind of is though.

> as the comment you replied to above points out, there are non-cryptographic functions with good avalanche properties which would satisfy that need without being collision-resistant.

No comment I replied to points anything near that. Your comment has basically no content, and the comment before that only asserts such existence without providing any guidance or evidence, linking to a page about the general concept of non-cryptographic hash function, which is utterly useless.

Not only that but avalanche properties do not matter at all for the use case: the hash is just a label for the sequence, it's fine if two similar sequences get similar hashes as long as those hashes are different. Some identifiers (like geninfo) are just centrally assigned integers.

[ivanbakel]

It's true that being collision-resistant is a strong enough property to make collisions unlikely, but it doesn't hold that collision resistance is a requirement for such a hash function.

What is the relevance of collision resistance in this case? Why do you say it's a primary requirement of a hash function here? Why isn't uniformity with a large enough image space enough? Given that there is no adversary trying to produce collisions of generated identifiers, why does it matter that collisions are hard to deliberately create, rather than simply unlikely to occur?