Encrypting/Hashing plain text passwords in databas

问题:

I\'ve inherited a web app that I\'ve just discovered stores over 300,000 usernames/passwords in plain text in a SQL Server database. I realize that this is a Very Bad Thing™.

Knowing that I\'ll have to update the login and password update processes to encrypt/decrypt, and with the smallest impact on the rest of the system, what would you recommend as the best way to remove the plain text passwords from the database?

Any help is appreciated.

Edit: Sorry if I was unclear, I meant to ask what would be your procedure to encrypt/hash the passwords, not specific encryption/hashing methods.

Should I just:

1. Make a backup of the DB
2. Update login/update password code
3. After hours, go through all records in the users table hashing the password and replacing each one
4. Test to ensure users can still login/update passwords

I guess my concern is more from the sheer number of users so I want to make sure I\'m doing this correctly.

回答1:

I would imagine you will have to add a column to the database for the encrypted password then run a batch job over all records which gets the current password, encrypts it (as others have mentiond a hash like md5 is pretty standard edit: but should not be used on its own - see other answers for good discussions), stores it in the new column and checks it all happened smoothly.

Then you will need to update your front-end to hash the user-entered password at login time and verify that vs the stored hash, rather than checking plaintext-vs-plaintext.

It would seem prudent to me to leave both columns in place for a little while to ensure that nothing hinky has gone on, before eventually removing the plaintext passwords all-together.

Don\'t forget also that anytime the password is acessed the code will have to change, such as password change / reminder requests. You will of course lose the ability to email out forgotten passwords, but this is no bad thing. You will have to use a password reset system instead.

Edit: One final point, you might want to consider avoiding the error I made on my first attempt at a test-bed secure login website:

When processing the user password, consider where the hashing takes place. In my case the hash was calculated by the PHP code running on the webserver, but the password was transmitted to the page from the user\'s machine in plaintext! This was ok(ish) in the environment I was working in, as it was inside an https system anyway (uni network). But, in the real world I imagine you would want to hash the password before it leaves the user system, using javascript etc. and then transmit the hash to your site.

回答2:

EDIT (2016): use Argon2, scrypt, bcrypt, or PBKDF2, in that order of preference. Use as large a slowdown factor as is feasible for your situation. Use a vetted existing implementation. Make sure you use a proper salt (although the libraries you\'re using should be making sure of this for you).

---

When you hash the passwords use DO NOT USE PLAIN MD5.

Use PBKDF2, which basically means using a random salt to prevent rainbow table attacks, and iterating (re-hashing) enough times to slow the hashing down - not so much that your application takes too long, but enough that an attacker brute-forcing a large number of different password will notice

From the document:

• Iterate at least 1000 times, preferably more - time your implementation to see how many iterations are feasible for you.
• 8 bytes (64 bits) of salt are sufficient, and the random doesn\'t need to be secure (the salt is unencrypted, we\'re not worried someone will guess it).
• A good way to apply the salt when hashing is to use HMAC with your favorite hash algorithm, using the password as the HMAC key and the salt as the text to hash (see this section of the document).

Example implementation in Python, using SHA-256 as the secure hash:

EDIT: as mentioned by Eli Collins this is not a PBKDF2 implementation. You should prefer implementations which stick to the standard, such as PassLib.

from hashlib import sha256
from hmac import HMAC
import random

def random_bytes(num_bytes):
  return \"\".join(chr(random.randrange(256)) for i in xrange(num_bytes))

def pbkdf_sha256(password, salt, iterations):
  result = password
  for i in xrange(iterations):
    result = HMAC(result, salt, sha256).digest() # use HMAC to apply the salt
  return result

NUM_ITERATIONS = 5000
def hash_password(plain_password):
  salt = random_bytes(8) # 64 bits

  hashed_password = pbkdf_sha256(plain_password, salt, NUM_ITERATIONS)

  # return the salt and hashed password, encoded in base64 and split with \",\"
  return salt.encode(\"base64\").strip() + \",\" + hashed_password.encode(\"base64\").strip()

def check_password(saved_password_entry, plain_password):
  salt, hashed_password = saved_password_entry.split(\",\")
  salt = salt.decode(\"base64\")
  hashed_password = hashed_password.decode(\"base64\")

  return hashed_password == pbkdf_sha256(plain_password, salt, NUM_ITERATIONS)

password_entry = hash_password(\"mysecret\")
print password_entry # will print, for example: 8Y1ZO8Y1pi4=,r7Acg5iRiZ/x4QwFLhPMjASESxesoIcdJRSDkqWYfaA=
check_password(password_entry, \"mysecret\") # returns True

回答3:

The basic strategy is to use a key derivation function to \"hash\" the password with some salt. The salt and the hash result are stored in the database. When a user inputs a password, the salt and their input are hashed in the same way and compared to the stored value. If they match, the user is authenticated.

The devil is in the details. First, a lot depends on the hash algorithm that is chosen. A key derivation algorithm like PBKDF2, based on a hash-based message authentication code, makes it \"computationally infeasible\" to find an input (in this case, a password) that will produce a given output (what an attacker has found in the database).

A pre-computed dictionary attack uses pre-computed index, or dictionary, from hash outputs to passwords. Hashing is slow (or it\'s supposed to be, anyway), so the attacker hashes all of the likely passwords once, and stores the result indexed in such a way that given a hash, he can lookup a corresponding password. This is a classic tradeoff of space for time. Since password lists can be huge, there are ways to tune the tradeoff (like rainbow tables), so that an attacker can give up a little speed to save a lot of space.

Pre-computation attacks are thwarted by using \"cryptographic salt\". This is some data that is hashed with the password. It doesn\'t need to be a secret, it just needs to be unpredictable for a given password. For each value of salt, an attacker would need a new dictionary. If you use one byte of salt, an attacker needs 256 copies of their dictionary, each generated with a different salt. First, he\'d use the salt to lookup the correct dictionary, then he\'d use the hash output to look up a usable password. But what if you add 4 bytes? Now he needs 4 billion copies of the the dictionary. By using a large enough salt, a dictionary attack is precluded. In practice, 8 to 16 bytes of data from a cryptographic quality random number generator makes a good salt.

With pre-computation off the table, an attacker has compute the hash on each attempt. How long it takes to find a password now depends entirely on how long it takes to hash a candidate. This time is increased by iteration of the hash function. The number iterations is generally a parameter of the key derivation function; today, a lot of mobile devices use 10,000 to 20,000 iterations, while a server might use 100,000 or more. (The bcrypt algorithm uses the term \"cost factor\", which is a logarithmic measure of the time required.)

回答4:

Follow Xan\'s advice of keeping the current password column around for a while so if things go bad, you can rollback quick-n-easy.

As far as encrypting your passwords:

• use a salt
• use a hash algorithm that\'s meant for passwords (ie., - it\'s slow)

See Thomas Ptacek\'s Enough With The Rainbow Tables: What You Need To Know About Secure Password Schemes for some details.

回答5:

I think you should do the following:

1. Create a new column called HASHED_PASSWORD or something similar.
2. Modify your code so that it checks for both columns.
3. Gradually migrate passwords from the non-hashed table to the hashed one. For example, when a user logs in, migrate his or her password automatically to the hashed column and remove the unhashed version. All newly registered users will have hashed passwords.
4. After hours, you can run a script which migrates n users a time
5. When you have no more unhashed passwords left, you can remove your old password column (you may not be able to do so, depends on the database you are using). Also, you can remove the code to handle the old passwords.
6. You\'re done!

回答6:

That was a problem of mine couple of weeks ago. We were deploying a large MIS project to 975 different geographical locations where our own user credential store will be used as an authenticator for different set of already implemented and in-use applications. We already provided both REST and SOAP based authentication service but customer insisted to be able to reach the user credential store from other applications with just a a DB connection to read-only view of related table or view. Sigh... (this highly coupled bad design decision is a subject of another question).

That forced us to sit and convert our salted and iteratively hashed password storage scheme to a specification and provide some different language implementations for easy integration.

We called it Fairly Secure Hashed Passwords or FSHP in short. Implemented it in Python, Ruby, PHP5 and released it to Public Domain. Available to be consumed, forked, flamed or spit on GitHub at http://github.com/bdd/fshp

FSHP is a salted, iteratively hashed password hashing implementation.

Design principle is similar with PBKDF1 specification in RFC 2898 (a.k.a: PKCS #5: Password-Based Cryptography Specification Version 2.0.) FSHP allows choosing the salt length, number of iterations and the underlying cryptographic hash function among SHA-1 and SHA-2 (256, 384, 512). Self defining meta prefix at the beginning of every output makes it portable while letting the consumer to choose its own password storage security baseline.

SECURITY:

Default FSHP1 uses 8 byte salts, with 4096 iterations of SHA-256 hashing. - 8 byte salt renders rainbow table attacks impractical by multiplying the required space with 2^64. - 4096 iterations causes brute force attacks to be fairly expensive. - There are no known attacks against SHA-256 to find collisions with a computational effort of fewer than 2^128 operations at the time of this release.

IMPLEMENTATIONS:

• Python: Tested with 2.3.5 (w/ hashlib), 2.5.1, 2.6.1
• Ruby : Tested with 1.8.6
• PHP5 : Tested with 5.2.6

Everyone is more than welcome to create missing language implementations or polish the current ones.

BASIC OPERATION (with Python):

>>> fsh = fshp.crypt(\'OrpheanBeholderScryDoubt\')
>>> print fsh
{FSHP1|8|4096}GVSUFDAjdh0vBosn1GUhzGLHP7BmkbCZVH/3TQqGIjADXpc+6NCg3g==
>>> fshp.validate(\'OrpheanBeholderScryDoubt\', fsh)
True

CUSTOMIZING THE CRYPT:

Let\'s weaken our password hashing scheme. - Decrease the salt length from default 8 to 2. - Decrease the iteration round from default 4096 to 10. - Select FSHP0 with SHA-1 as the underlying hash algorithm.

>>> fsh = fshp.crypt(\'ExecuteOrder66\', saltlen=2, rounds=10, variant=0)
>>> print fsh
{FSHP0|2|10}Nge7yRT/vueEGVFPIxcDjiaHQGFQaQ==

回答7:

As the others mentioned, you don\'t want to decrypt if you can help it. Standard best practice is to encrypt using a one-way hash, and then when the user logs in hash their password to compare it.

Otherwise you\'ll have to use a strong encryption to encrypt and then decrypt. I\'d only recommend this if the political reasons are strong (for example, your users are used to being able to call the help desk to retrieve their password, and you have strong pressure from the top not to change that). In that case, I\'d start with encryption and then start building a business case to move to hashing.

回答8:

For authentication purposes you should avoid storing the passwords using reversible encryption, i.e. you should only store the password hash and check the hash of the user-supplied password against the hash you have stored. However, that approach has a drawback: it\'s vulnerable to rainbow table attacks, should an attacker get hold of your password store database.

What you should do is store the hashes of a pre-chosen (and secret) salt value + the password. I.e., concatenate the salt and the password, hash the result, and store this hash. When authenticating, do the same - concatenate your salt value and the user-supplied password, hash, then check for equality. This makes rainbow table attacks unfeasible.

Of course, if the user send passwords across the network (for example, if you\'re working on a web or client-server application), then you should not send the password in clear text across, so instead of storing hash(salt + password) you should store and check against hash(salt + hash(password)), and have your client pre-hash the user-supplied password and send that one across the network. This protects your user\'s password as well, should the user (as many do) re-use the same password for multiple purposes.

回答9:

• Encrypt using something like MD5, encode it as a hex string
• You need a salt; in your case, the username can be used as the salt (it has to be unique, the username should be the most unique value available ;-)
• use the old password field to store the MD5, but tag the MD5 (i.e.g \"MD5:687A878....\") so that old (plain text) and new (MD5) passwords can co-exist
• change the login procedure to verify against the MD5 if there is an MD5, and against the plain password otherwise
• change the \"change password\" and \"new user\" functions to create MD5\'ed passwords only
• now you can run the conversion batch job, which might take as long as needed
• after the conversion has been run, remove the legacy-support

回答10:

Step 1: Add encrypted field to database

Step 2: Change code so that when password is changed, it updates both fields but logging in still uses old field.

Step 3: Run script to populate all the new fields.

Step 4: Change code so that logging in uses new field and changing passwords stops updating old field.

Step 5: Remove unencrypted passwords from database.

This should allow you to accomplish the changeover without interruption to the end user.

Also: Something I would do is name the new database field something that is completely unrelated to password like \"LastSessionID\" or something similarly boring. Then instead of removing the password field, just populate with hashes of random data. Then, if your database ever gets compromised, they can spend all the time they want trying to decrypt the \"password\" field.

This may not actually accomplish anything, but it\'s fun thinking about someone sitting there trying to figure out worthless information

回答11:

As with all security decisions, there are tradeoffs. If you hash the password, which is probably your easiest move, you can\'t offer a password retrieval function that returns the original password, nor can your staff look up a person\'s password in order to access their account.

You can use symmetric encryption, which has its own security drawbacks. (If your server is compromised, the symmetric encryption key may be compromised also).

You can use public-key encryption, and run password retrieval/customer service on a separate machine which stores the private key in isolation from the web application. This is the most secure, but requires a two-machine architecture, and probably a firewall in between.

回答12:

I\'m not a security expert, but i htink the current recommendation is to use bcrypt/blowfish or a SHA-2 variant, not MD5 / SHA1.

Probably you need to think in terms of a full security audit, too

回答13:

MD5 and SHA1 have shown a bit of weakness (two words can result in the same hash) so using SHA256-SHA512 / iterative hashes is recommended to hash the password.

I would write a small program in the language that the application is written in that goes and generates a random salt that is unique for each user and a hash of the password. The reason I tend to use the same language as the verification is that different crypto libraries can do things slightly differently (i.e. padding) so using the same library to generate the hash and verify it eliminates that risk. This application could also then verify the login after the table has been updated if you want as it knows the plain text password still.

1. Don\'t use MD5/SHA1
2. Generate a good random salt (many crypto libraries have a salt generator)
3. An iterative hash algorithm as orip recommended
4. Ensure that the passwords are not transmitted in plain text over the wire

回答14:

I would like to suggest one improvement to the great python example posted by Orip. I would redefine the random_bytes function to be:

def random_bytes(num_bytes):
    return os.urandom(num_bytes)

Of course, you would have to import the os module. The os.urandom function provides a random sequence of bytes that can be safely used in cryptographic applications. See the reference help of this function for further details.

回答15:

To hash the password you can use the HashBytes function. Returns a varbinary, so you\'d have to create a new column and then delete the old varchar one.

Like

ALTER TABLE users ADD COLUMN hashedPassword varbinary(max);
ALTER TABLE users ADD COLUMN salt char(10);
--Generate random salts and update the column, after that
UPDATE users SET hashedPassword = HashBytes(\'SHA1\',salt + \'|\' + password);

Then you modify the code to validate the password, using a query like

SELECT count(*) from users WHERE hashedPassword = 
HashBytes(\'SHA1\',salt + \'|\' + <password>)

where <password> is the value entered by the user.

回答16:

hash them with md5. that\'s what is usually done with passwords.