Hcfyk

I know that MD5 should not be used for password hashing, and that it also should not be used for integrity checking of documents. There are way too many sources citing MD5 preimaging attacks and MD5s low computation time.

There is no published preimage attack on MD5 that is cheaper than a generic attack on any 128-bit hash function. But you shouldn't rely on that alone when making security decisions, because cryptography is tricky and adversaries are clever and resourceful and can find ways around it!

1. Identifying malicious files, such as when Linux Mint's download servers were compromised and an ISO file was replaced by a malicious one; in this case you want to be sure that your file doesn't match; collision attacks aren't a vector here.

The question of whether to publish known-good vs. known-bad hashes after a compromise is addressed elsewhere—in brief, there's not much that publishing known-bad hashes accomplishes, and according to the citation, Linux Mint published known-good, not known-bad, hashes. So what security do you get from known-good MD5 hashes?

There are two issues here:

1. 
   

   If you got the MD5 hash from the same source as the ISO image, there's nothing that would prevent an adversary from replacing both the MD5 hash and the ISO image.

   
   
   

   To prevent this, you and the Linux Mint curators need two channels: one for the hashes which can't be compromised (but need only have very low bandwidth), and another for the ISO image (which needs high bandwidth) on which you can then use the MD5 hash in an attempt to detect compromise.

   
   
   

   There's another way to prevent this: Instead of using the uncompromised channel for the hash of every ISO image over and over again as time goes on—which means more and more opportunities for an attacker to subvert it—use it once initially for a public key, which is then used to sign the ISO images; then there's only one opportunity for an attacker to subvert the public key channel.

   
   


2. 
   

   Collision attacks may still be a vector in cases like this. Consider the following scenario:

   
   
   
   
   
   • I am an evil developer. I write two software packages, whose distributions collide under MD5. One of the packages is benign and will survive review and audit. The other one will surreptitiously replace your family photo album by erotic photographs of sushi.
   
   
   • The Linux Mint curators carefully scrutinize and audit everything they publish in their package repository and publish the MD5 hashes of what they have audited in a public place that I can't compromise.
   
   
   • The Linux Mint curators cavalierly administer the package distributions in their package repository, under the false impression that the published MD5 hashes will protect users.
   
   
   
   

   In this scenario, I can replace the benign package by the erotic sushi package, pass the MD5 verification with flying colors, and give you a nasty—and luscious—surprise when you try to look up photos of that old hiking trip you took your kids on.

   
   

2. Finding duplicate files. By MD5-summing all files in a directory structure it's easy to find identical hashes. The seemingly identical files can then be compared in full to check if they are really identical. Using SHA512 would make the process slower, and since we compare files in full anyway there is no risk in a potential false positive from MD5.

When I put my benign software package and my erotic sushi package, which collide under MD5, in your directory, your duplicate-detection script will initially think they are duplicates. In this case, you absolutely must compare the files in full. But there are much better ways to do this!

• If you use SHA-512, you can safely skip the comparison step. Same if you use BLAKE2b, which can be even faster than MD5.




• You could even use MD5 safely for this if you use it as HMAC-MD5 under a uniform random key, and safely skip the comparison step. HMAC-MD5 does not seem to be broken, as a pseudorandom function family—so it's probably fine for security, up to the birthday bound, but there are better faster PRFs like keyed BLAKE2 that won't raise any auditors' eyebrows.




• Even better, you can choose a random key and hash the files with a universal hash under the key, like Poly1305. This is many times faster than MD5 or BLAKE2b, and the probability of a collision between any two files is less than $1/2^{100}$, so the probability of collision among $n$ files is less than $binom n 2 2^{-100}$ and thus you can still safely skip the comparison step until you have quadrillions of files.




• You could also just use a cheap checksum like a CRC with a fixed polynomial. This will be the fastest of the options—far and away faster than MD5—but unlike the previous options you still absolutely must compare the files in full.

So, is MD5 safe for finding candidate duplicates to verify, if you subsequently compare the files bit by bit in full? Yes. So is the constant zero function.

(In a way, this would be creating a rainbow table where all the files are the dictionary)

This is not a rainbow table. A rainbow table is a specific technique for precomputing a random walk over a space of, say, passwords, via, say, MD5 hashes, in a way that saves effort trying to find MD5 preimages for hashes that aren't necessarily in your table in the first place, or doing it in parallel to speed up a multi-target search. It is not simply a list of precomputed hashes on a dictionary of inputs.

(The blog post by tptacek that you cited, and the blog post by Jeff Atwood that it was a response to, are both confused about what rainbow tables are.)

When the password scheme article states that "MD5 is fast", it clearly refers to the problem that hashing MD5 is too cheap when it comes to hashing a large amount of passwords to find the reverse of a hash. But what does it mean when it says that "[MD5 is] too slow to use as a general purpose hash"? Are there faster standardized hashes to compare files, that still have a reasonably low chance of collision?

I don't know what tptacek meant—you could email and ask—but if I had to guess, I would guess this meant it's awfully slow for things like hash tables, where you would truncate MD5 to a few bits to determine an index into an array of buckets or an open-addressing array.

But what does it mean when it says that "[MD5 is] too slow to use as a general purpose hash"? Are there faster standardized hashes to compare files, that still have a reasonably low chance of collision?

BLAKE2 is faster than MD5 and currently known to provide 64-bit collision resistence when truncated to the same size as MD5 (compare ~30 of that of MD5).

There's not a compelling reason to use MD5; however, there are some embedded systems with a MD5 core that was used as a stream verifier. In those systems, MD5 is still used. They are moving to BLAKE2 because it's smaller in silicon, and it has the benefit of being faster than MD5 in general.

The reason that MD5 started fall out of favor with hardware people was that the word reordering of the MD5 message expansions seems to be simple, but actually
they require a lot of circuits for demultiplexing and interconnect, and the hardware efficiencies are greatly degraded compared to BLAKE. In contrast, the message expansion blocks for BLAKE algorithms can be efficiently implemented as simple feedback shift registers.

The BLAKE team did a nice job of making it work well on silicon and in instructions.

edit: SHA-1, SHA-2, etc also look pretty nice in circuits.

I know that MD5 should not be used for password hashing

Indeed. However, that's about the direct applicability of MD5 to a password or to use it with just a password and salt. In that case MD5 is less secure than a dedicated password hash with a work factor, at least for common passwords and pass phrases.

However, the use of MD5 within a PRF (HMAC) and within a password hash is still OK as it relies on pre-image resistance for security, rather than collision resistance.

I'd rather not bet that MD5 stays secure for pre-image resistance though. Attacks only get better and although I don't see any progress on breaking MD5's pre-image resistance, I would not rule it out either.

Identifying malicious files, such as when Linux Mint's download servers were compromised and an ISO file was replaced by a malicious one; in this case you want to be sure that your file doesn't match; collision attacks aren't a vector here.

MD5 is still secure to check hashes from another server as long as hackers cannot alter the input of the MD5 hash. However, for something like a full ISO, I'd say that they would have plenty of opportunity bringing in binary files that seem innocent with regard to contents, while they alter the intermediate state of MD5 vulnerable to collision attacks.

That was not the attack that you referred to; in that case the MD5 hash on the official server was different from the one calculated over the ISO image.

But attack on file distribution do rely on collision resistance, and this kind of use case can definitely be attacked. It would probably not be all that easy (correctly lining up the binary data required for the attack at the start of the ISO and such), but it is a vector of attack none-the-less.

The same goes for SHA-1 in Git by the way. Not easy to breach, but far from impossible, whatever Linus says.

Finding duplicate files. By MD5-summing all files in a directory structure it's easy to find identical hashes. The seemingly identical files can then be compared in full to check if they are really identical. Using SHA512 would make the process slower, and since we compare files in full anyway there is no risk in a potential false positive from MD5. (In a way, this would be creating a rainbow table where all the files are the dictionary)

Sure, if there is no possibility of attack or if you compare the files fully anyway then MD5 is fine.

However, if there is a vector of attack then you need to perform a full binary compare even after the hash has matched. Otherwise an attacker could make you retrieve the wrong deduplicated file. If you'd use a cryptographically strong hash then you would not have to perform the full file comparison at all. As others have noticed, a 256-512 bit hash is a lot easier to handle than performing a full file compare when storing. Passing over the file twice is not very fast either; all the speed advantages of MD5 are very likely nullified by the I/O required.

Besides that, if you would reference it using the hash then there is no comparison to be made; you would only have a single file (this is about deduplication after all).