# AES: Why is it a good practice to use only the first 16 bytes of a hash for encryption?

I'd like to encrypt Text with AES/CTR and a password defined by the user in java. I already checked the internet (and stackoverflow) for answers. The most used version is to hash the user password with sha1 and take only the first 16 bytes.

But I don't think this can be a good practice.

1. sha1 is weak
2. taking only the first 16 bytes makes the hash also weak and rise the chance for a collision (even with sha-256)

Is this really the best practice? Why? How can I do things better?

Some links to the articles I mentioned:

• They are not good sources. Anyway I will call this question as dupe of this and this – kelalaka Apr 4 at 18:28
• Nowadays you should probably use HKDF with an appropriate hash. – jww Apr 5 at 2:42
• Collision is irrelevant. The chance of 128-bit hash (thus key) collision for different passwords is very low unless you have at least thousands of times more users than exist on Earth (about 2^44) and anyway there's no harm as long as you do not also reuse the same nonce for different data. If the (96-bit) nonce is indepedently random that chance is infinitesimal; if the nonce is systematic (e.g. counter) or synthetic (SIV) there is zero chance. The real and serious danger is using a fast hash on password input, as correctly answered by Ella. – dave_thompson_085 Apr 5 at 15:50

Why is it a good practice to use only the first 16 bytes of a hash for encryption?

As you noted, it isn't.

But, the problem is not with the "16 bytes" part of the statement, or the concern for collisions. The problem is with the "hash" part.

## 16 bytes

As stated in one of the links you shared, AES only uses key sizes of 128, 192, and 256 bits (or 16, 24, and 32 bytes, respectively). So the key must be one of these sizes, because AES simply does not support other key sizes.

Trying to use a larger key could have a variety of possible outcomes depending on what the implementation chooses to do. It might raise an exception, or continue silently while only using the first N bits of the supplied key.

## Hashing a password to use as an encryption key

Using a hash function such as MD5, SHA1, SHA2, SHA3, blake2, etc, would all be bad practice. The first two are obvious: MD5 and SHA1 are known to be weak in general.

But even using a strong cryptographic hash like SHA3 or blake2 would also be bad, because they were not designed to solve the problem of deriving a key from a password. Use of a cryptographic hash function is involved in this process, but it is not the entirety of it.

Good practice would be to use a dedicated key derivation function such as Argon2 that was designed to solve this problem. If your library doesn't support Argon2 but supports scrypt, bcrypt or PBKDF2, any of these three is also a reasonable choice.

## Why/How

A normal hash function is designed to be fast and require little space.

A hash function designed for use on passwords is quite the opposite: it is a slow function that requires lots of memory access, in an attempt to try and optimize the function towards what a consumer CPU is good at, and minimize the potential for optimization with special hardware. Specialized hardware is usable by an attacker, but a legitimate user is limited to a commodity CPU; The goal is to try and use a function that cannot take advantage of special hardware to the extent possible.

Details about the hows and whys of password hashing are listed in this paper and quoted below (with minor modifications, e.g. removing citations and modified formatting):

Cryptographic Security: The scheme should be cryptographically secure and as such possess the following properties:

• 1) Preimage resistance
• 2) Second preimage resistance
• 3) collision resistance.

In addition it should avoid other cryptographic weaknesses such as those present in (some)Merkle-Damgård constructions(e.g. length extension attacks, partial message collisions, etc)

Defense against lookup table /TMTOAttacks:

• The scheme should aim to make TMTO attacks that allow for precomputed lookup table generation, such as Rainbow Tables, infeasible

Defense against CPU-optimized 'crackers':

• The scheme should be ‘CPU-hard’, that is, it should require significant amounts of CPU processing in a manner that cannot be optimized away through either software or hardware. As such, cracking-optimized (multi-core) CPU software implementations (eg. written in assembly, testing multiple input sets in parallel) should offer only minimal speed-up improvements compared to those intended for validation (“slower for attackers, faster for defenders”).

Defense against hardware-optimized 'crackers':

• The scheme should be 'memory-hard', that is, it should significant amounts of RAM capacity in a manner that cannot be optimized away through eg. TMTO attacks. As such cracking-optimized ASIC, FPGA and GPU implementations should offer only minimal speed up improvements (eg. in terms of time-area product) compared to those intended for validation. As noted by Aumasson one of the main scheme design challenges is ensuring minimized efficiency on GPUs, FPGAs and ASICs (in order to minimize benefits of cracking-optimized implementations) and maximized efficiency on general-purpose CPUs (in order to maintain regular use efficiency).

Defense against side-channel attacks:

• Depending on the use-case (eg. for key derivation or authentication to a device seeking to protect against modification by the device owner) side-channel attacks might be a relevant avenue of attack. Password hashing schemes should aim to offer side-channel resilience. With regards to password hashing scheme security we will focus on security versus the cache-timing type of side-channel attacks given the existence of such attacks against the commonly used scrypt scheme. The second category of side-channel attacks we will take into consideration are so-called Garbage Collector Attacks (GCAs). GCAs have been discussed in literature as an instance of a 'memory leak' attack relevant to password hashing scheme security. GCAs consist of a scenario where an attacker has access to a target machine's internal memory either after termination of the hashing scheme or at some point where the password itself is still present in memory (the so-called WeakGCA variant)...
• Nitpick: bcrypt is advertised as a password storage and verification function, not so much a key derivation function, and implementations routinely have APIs to match that (e.g., outputting text encoded output, providing an enroll/verify API instead of a hash API, That is not to claim that bcrypt couldn't be used as you suggest, but there are potential practical pitfalls. See, e.g., this article. – Luis Casillas Apr 4 at 22:00
• @LuisCasillas just a note: I actually didn't list bcrypt; that was inserted to my answer by Gilles via an edit... – Ella Rose Apr 4 at 22:35
• @LuisCasillas Argon2 was also the winner of the password hashing competition, not the password-based KDF competition. Is there any reason to believe that Argon2 is good for PBKDF that doesn't also apply to bcrypt? – Gilles Apr 4 at 23:46
• @Gilles The PHC call for submissions had a requirement that the outputs look random, which to my mind implies such suitability. But really, the biggest pitfalls I'm thinking of here aren't conceptually deep; they come down to bcrypt implementations doing stuff like producing ASCII output like "$2a$" + cost + "\$" + base64(salt + hash). At the level of (un)sophistication we're dealing with in this question I worry somebody might literally use the ASCII string's bytes. – Luis Casillas Apr 5 at 9:46
• @firendlyQuestion: Yes, provided you use an Argon2 implementation or API that produces binary output. The tricky detail here is that some implementations of such functions are coded to be friendly to callers who are using them for password storage instead of key derivation. – Luis Casillas Apr 5 at 9:48