I'm developing a REST API that requires authentication. Because the authentication itself occurs via an external webservice over HTTP, I reasoned that we would dispense tokens to avoid repeatedly calling the authentication service. Which brings me neatly to my first question:
Is this really any better than just requiring clients to use HTTP Basic Auth on each request and caching calls to the authentication service server-side?
The Basic Auth solution has the advantage of not requiring a full round-trip to the server before requests for content can begin. Tokens can potentially be more flexible in scope (i.e. only grant rights to particular resources or actions), but that seems more appropriate to the OAuth context than my simpler use case.
Currently tokens are acquired like this:
curl -X POST localhost/token --data "api_key=81169d80...
&verifier=2f5ae51a...
×tamp=1234567
&user=foo
&pass=bar"
The api_key, timestamp and verifier are required by all requests. The "verifier" is returned by:
sha1(timestamp + api_key + shared_secret)
My intention is to only allow calls from known parties, and to prevent calls from being reused verbatim.
Is this good enough? Underkill? Overkill?
With a token in hand, clients can acquire resources:
curl localhost/posts?api_key=81169d80...
&verifier=81169d80...
&token=9fUyas64...
×tamp=1234567
For the simplest call possible, this seems kind of horribly verbose. Considering the shared_secret will wind up being embedded in (at minimum) an iOS application, from which I would assume it can be extracted, is this even offering anything beyond a false sense of security?
Let me seperate up everything and solve approach each problem in isolation:
Authentication
For authentication, baseauth has the advantage that it is a mature solution on the protocol level. This means a lot of "might crop up later" problems are already solved for you. For example, with BaseAuth, user agents know the password is a password so they don't cache it.
Auth server load
If you dispense a token to the user instead of caching the authentication on your server, you are still doing the same thing: Caching authentication information. The only difference is that you are turning the responsibility for the caching to the user. This seems like unnecessary labor for the user with no gains, so I recommend to handle this transparently on your server as you suggested.
Transmission Security
If can use an SSL connection, that's all there is to it, the connection is secure*. To prevent accidental multiple execution, you can filter multiple urls or ask users to include a random component ("nonce") in the URL.
url = username:key@myhost.com/api/call/nonce
If that is not possible, and the transmitted information is not secret, I recommend securing the request with a hash, as you suggested in the token approach. Since the hash provides the security, you could instruct your users to provide the hash as the baseauth password. For improved robustness, I recommend using a random string instead of the timestamp as a "nonce" to prevent replay attacks (two legit requests could be made during the same second). Instead of providing seperate "shared secret" and "api key" fields, you can simply use the api key as shared secret, and then use a salt that doesn't change to prevent rainbow table attacks. The username field seems like a good place to put the nonce too, since it is part of the auth. So now you have a clean call like this:
nonce = generate_secure_password(length: 16);
one_time_key = nonce + '-' + sha1(nonce+salt+shared_key);
url = username:one_time_key@myhost.com/api/call
It is true that this is a bit laborious. This is because you aren't using a protocol level solution (like SSL). So it might be a good idea to provide some kind of SDK to users so at least they don't have to go through it themselves. If you need to do it this way, I find the security level appropriate (just-right-kill).
Secure secret storage
It depends who you are trying to thwart. If you are preventing people with access to the user's phone from using your REST service in the user's name, then it would be a good idea to find some kind of keyring API on the target OS and have the SDK (or the implementor) store the key there. If that's not possible, you can at least make it a bit harder to get the secret by encrypting it, and storing the encrypted data and the encryption key in seperate places.
If you are trying to keep other software vendors from getting your API key to prevent the development of alternate clients, only the encrypt-and-store-seperately approach almost works. This is whitebox crypto, and to date, no one has come up with a truly secure solution to problems of this class. The least you can do is still issue a single key for each user so you can ban abused keys.
(*) EDIT: SSL connections should no longer be considered secure without taking additional steps to verify them.
A pure RESTful API should use the underlying protocol standard features:
For HTTP, the RESTful API should comply with existing HTTP standard headers. Adding a new HTTP header violates the REST principles. Do not re-invent the wheel, use all the standard features in HTTP/1.1 standards - including status response codes, headers, and so on. RESTFul web services should leverage and rely upon the HTTP standards.
RESTful services MUST be STATELESS. Any tricks, such as token based authentication that attempts to remember the state of previous REST requests on the server violates the REST principles. Again, this is a MUST; that is, if you web server saves any request/response context related information on the server in attempt to establish any sort of session on the server, then your web service is NOT Stateless. And if it is NOT stateless it is NOT RESTFul.
Bottom-line: For authentication/authorization purposes you should use HTTP standard authorization header. That is, you should add the HTTP authorization / authentication header in each subsequent request that needs to be authenticated. The REST API should follow the HTTP Authentication Scheme standards.The specifics of how this header should be formatted are defined in the RFC 2616 HTTP 1.1 standards – section 14.8 Authorization of RFC 2616, and in the RFC 2617 HTTP Authentication: Basic and Digest Access Authentication.
I have developed a RESTful service for the Cisco Prime Performance Manager application. Search Google for the REST API document that I wrote for that application for more details about RESTFul API compliance here. In that implementation, I have chosen to use HTTP "Basic" Authorization scheme. - check out version 1.5 or above of that REST API document, and search for authorization in the document.
In the web a stateful protocol is based on having a temporary token that is exchanged between a browser and a server (via cookie header or URI rewriting) on every request. That token is usually created on the server end, and it is a piece of opaque data that has a certain time-to-live, and it has the sole purpose of identifying a specific web user agent. That is, the token is temporary, and becomes a STATE that the web server has to maintain on behalf of a client user agent during the duration of that conversation. Therefore, the communication using a token in this way is STATEFUL. And if the conversation between client and server is STATEFUL it is not RESTful.
The username/password (sent on the Authorization header) is usually persisted on the database with the intent of identifying a user. Sometimes the user could mean another application; however, the username/password is NEVER intended to identify a specific web client user agent. The conversation between a web agent and server based on using the username/password in the Authorization header (following the HTTP Basic Authorization) is STATELESS because the web server front-end is not creating or maintaining any STATE information whatsoever on behalf of a specific web client user agent. And based on my understanding of REST, the protocol states clearly that the conversation between clients and server should be STATELESS. Therefore, if we want to have a true RESTful service we should use username/password (Refer to RFC mentioned in my previous post) in the Authorization header for every single call, NOT a sension kind of token (e.g. Session tokens created in web servers, OAuth tokens created in authorization servers, and so on).
I understand that several called REST providers are using tokens like OAuth1 or OAuth2 accept-tokens to be be passed as "Authorization: Bearer " in HTTP headers. However, it appears to me that using those tokens for RESTful services would violate the true STATELESS meaning that REST embraces; because those tokens are temporary piece of data created/maintained on the server side to identify a specific web client user agent for the valid duration of a that web client/server conversation. Therefore, any service that is using those OAuth1/2 tokens should not be called REST if we want to stick to the TRUE meaning of a STATELESS protocol.
Rubens
