[kpdemetriou]

Absolutely, but NIST ultimately choose the winners, giving them the option to pick (non-obviously) weak/weaker algorithms. Historically only the winners are adopted. Look at the AES competition - how often do you see Serpent being mentioned, despite it having a larger security margin than Rijndael by most accounts?

[lucb1e]

> Historically only the winners are adopted. Look at the AES competition

Often, yes. But also consider the SHA-3 competition.

BLAKE2 seems more widely used than what was chosen for SHA-3 (Keccak). What was submitted for the SHA-3 competition was BLAKE1 (it didn't have a number back then but I think this is clearer) so it's not like NIST said that Keccak is better than BLAKE2, they only said it's better than BLAKE1 (per their requirements, which are unlikely to align with your requirements because of the heavy weighing of speed-in-hardware), but still this is an example of a widely used algorithm that is not standardized.

> how often do you see Serpent being mentioned, despite it having a larger security margin than Rijndael

The goal of an encryption algorithm is not only to be secure. Sure, that has to be a given: nobody is going to use a broken algorithm when given a choice. But when you have two secure options, the more efficient one is the one to choose. You could use a 32k RSA key just to be sure, or a 4k RSA key which (to the best of my knowledge) everyone considers safe until quantum. (After quantum, you need something like a 1TB key, as djb humorously proposed.)

Wikipedia article on Serpent: "The 32 rounds mean that Serpent has a higher security margin than Rijndael; however, Rijndael with 10 rounds is faster and easier to implement for small blocks."

I don't know that nobody talks about Serpent solely because it was not chosen as winner. It may just be that Rijndael with 256-bit keys is universally considered secure and is more efficient at doing its job.

[kpdemetriou]

Re: BLAKE2, I'm not sure it's fair to say that BLAKE2 is more widely used overall. But I do agree BLAKE2 is a bit of an outlier in terms of adoption. I think part of the reason is that SHA2 remains the go-to option, else I'd expect the ecosystem to consolidate around SHA3.

Re: Serpent, there are many things to unpack here but, in summary, you don't know a priori how large of a security margin you need (given the primary function of a cipher, you want to pick the conservative option), efficiency concerns become much less relevant with hardware-accelerated implementations and years of Moore's law performance uplifts, low-power devices can take advantage of much lighter algorithms than Rijndael OR Serpent, ease of implementation does not equal ease of correct/secure implementation vis-a-vis side channel attacks, and certainly if Serpent was chosen you wouldn't see Rijndael talked about much.

[pclmulqdq]

Blake2 also uses a very SHA2-like construction (a HAIFA construction, which is based on Merkle-Damgard). I believe this was the main reason SHA3 was chosen to be something completely different (a sponge construction). If SHA2 was found to be insecure, Blake2 would be at more risk of also being broken than Keccak.

Speculatively, if SHA2 is broken without breaking Merkle-Damgard hashes in general, Blake2/3 could well become SHA4.

[lucb1e]

> I'm not sure it's fair to say that BLAKE2 is more widely used overall

Ah, maybe my experience is biased then. I keep coming across BLAKE2 implementations, but rarely hear so much as people considering to use SHA-3 somewhere. If anyone has actual numbers on this, that would be interesting.

It would be good if SHA-3 is being used because then chip makers have a reason to bake it into their hardware, which is exactly where the biggest gain over SHA-2 is expected. If that happens, and all else being equal (no cracks appear in Keccak), I'd be surprised if BLAKE2 remains as popular!

> you don't know a priori how large of a security margin you need

True, so this can be argued to be an educated guess at first. But then confidence increases over time. It seems to be expected that, more than 20 years later, people aren't considering Serpent anymore. Is it because it wasn't chosen as AES? Certainly partially, but BLAKE2 (I'll admit it does seem like an outlier) likely will still be talked about in the future so standardization is not the only factor.

I didn't see actual benchmarks, but Serpent sounds at least three times slower than Rijndael for, by now, no tangible benefit. What would be interesting is if there were AES competitors that are also fully unbroken and are more efficient than Rijndael, or easier to implement, etc.

[zahllos]

> I didn't see actual benchmarks, but Serpent sounds at least three times slower than Rijndael

You can read the original report on the candidates: https://nvlpubs.nist.gov/nistpubs/jres/106/3/j63nec.pdf

To save you some time, one software evaluation is on page 531. Serpent performs worse in software than Rijndael (AES) just about everywhere (note they use categories rather than precise metrics, but you can dig into that if you want). By contrast, one hardware evaluation is on page 539. Serpent has the highest throughput for the lowest "area" i.e. required. These results repeat on p541.

So it depends on whether you are comparing a software or hardware implementation.

[EGreg]

I wish chip makers would bake the elliptic curve used in Bitcoin and Ethereum (secp256k) as well, instead of the entire industry coalescing around secp256r, which many suspect was somehow weaker (since its parameters are some weird large number X instead of a hard-to-game number like 15, leading some to believe that the first X-1 candidates were tried and X was found to be weaker).

The real reason I would have liked that to be the case is so that one could use WebAuthn and Subtle Web Crypto to sign things and have the blockchain natively verify the signature.

As it is, I am hoping that EVM will roll out a precompiled signature verifier for secp256r, which is on the roadmap — they say!

[arjvik]

There are a few different on-chain implementations of secp256r1 signature verification for use with passkeys, my favorite of which is demoed at https://p256.alembic.tech

Work is also being done on SNARK-based cross-curve signature verification

But I fully agree, especially with the growing popularity of account abstraction, the EVM desperately needs a secp256r1 precompile!

[nextaccountic]

> BLAKE2 seems more widely used than what was chosen for SHA-3 (Keccak)

There's also BLAKE3 and it is amazing. How is its adoption going?

https://github.com/BLAKE3-team/BLAKE3

[oconnor663]

It's been a good week! https://github.com/openzfs/zfs/releases/tag/zfs-2.2.0

[adastra22]

SHA-3 is vastly more widely used than BLAKE.

[zahllos]

I fully admit to having a weak spot for Serpent - it is self-bitslicing (see the submission package or the linux kernel tree), which in hindsight makes constant time software easier to write, and it was faster in hardware even when measured at the time, which is where we have ended up putting AES anyway (e.g. AES-NI etc).

BUT. On security margins, you could argue the Serpent designers were too conservative: https://eprint.iacr.org/2019/1492 It is also true that cryptanalytic attacks appear to fare slightly better against AES than Serpent. What does this mean? A brute force attack has the same number of operations as the claimed security level, say, 2^128 for 128-bit. An attack is something better than this: fewer operations. All of the attacks we know about achieve slightly less than this security level - which is nonetheless still impossible to do - but that comes at a cost: they need an infeasible amount of memory. In terms of numbers: 9000 TB to reduce 2^128 to 2^126 against full-round AES according to a quick check of wikipedia. For reference, the lightweight crypto competition considered 2^112 to be sufficient margin. 2^126 is still impossible.

In practice, the difference between Serpent and AES in terms of cryptanalytic security is meaningless. It is not an example of NIST picking a weaker algorithm deliberately, or I would argue, even unintentionally. It (AES) was faster when implemented in software for the 32-bit world that seemed to be the PC market at the time.

[kpdemetriou]

Implemented correctly, I agree the difference in security margin may not be too important. Otherwise, Serpent is more resistant to timing attacks. Weaknesses in implementation are as important as weaknesses in design.

Regardless, the comparison wasn't intended to argue for a meaningful difference in security margin, but to show that that the winner of the competition, well, wins (in adoption).

[bsder]

> BUT. On security margins, you could argue the Serpent designers were too conservative: https://eprint.iacr.org/2019/1492

Thanks for digging that paper out again. It is really telling that AES only gets a bit of a bump (10-30%) while the other ones gain like 2x or more.

I was about to comment that the competitors to AES were definitely too conservative, and it bit them because of how much slower it made them in software and larger in hardware.

[kevin_thibedeau]

Blowfish has a continuing existence as the basis for bcrypt.

[tptacek]

It works as a password hash for reasons having in part to do with why it isn’t a great general purpose cipher.

[bonzini]

Can you expand, or link to an explanation?

[crest]

The Blowfish key-schedule algorithm is equivalent to encrypting 4kB of data with it. This isn't a problem for some use-cases (e.g. transferring a large file over HTTPS), but terrible for others e.g. a encrypting lots of short messages using different keys without being able to cache > 30x larger result of the key-schedule result. To make it worse the cipher uses four large 256 x 32bit S-boxes with data (key and plaintext) dependent indexes making it very hard to implement fast without adding a timing side-channel on anything more complex than a simple microcontroller. It also does very little computation per memory access. Blowfish is a fast cipher on a very simple 32bit CPU with >= 4kiB of fast memory, but modern CPUs offer a lot more compute throughput than memory throughput. There is also very little opportunity to exploit for even the most expensive OoO CPUs because almost every depends on a data dependent memory access within a few instructions. For these reasons it's also expensive and relatively slow to implement in hardware.

Almost all of these downsides are helpful for password has validation function like bcrypt() because there is nothing to an attacker can to guess much faster than a desktop CPU.

Blowfish was a good cipher at a time when CPUs lacked dedicated trustworthy crypto engines, wide and deep OoO execution capability, and packed-SIMD support. AES and SHA1/2 are commonly implemented in hardware on modern server, desktop and mobile CPUs. Where hardware offloading isn't available ciphers can take advantage of OoO and SIMD to perform vastly more useful work per cycle than stalling on memory accesses.

[aidenn0]

Blowfish has an unusually slow key-setup phase. Slowness is an advantage for password hashes, since it makes offline attacks harder.