The problem is a common one when computing hashes for any data format where flexibility is allowed. To solve this, you need to canonicalize the representation.
For example, the OAuth1.0a protocol, which is used by Twitter and other services for authentication, requires a secure hash of the request message. To compute the hash, OAuth1.0a says you need to first alphabetize the fields, separate them by newlines, remove the field names (which are well known), and use blank lines for empty values. The signature or hash is computed on the result of that canonicalization.
XML DSIG works the same way – you need to canonicalize the XML before signing it. There is a proposed W3 standard covering this, because it’s such a fundamental requirement for signing. Some people call it c14n.
I don’t know of a canonicalization standard for json. It’s worth researching.
If there isn’t one, you can certainly establish a convention for your particular application usage. A reasonable start might be:
- lexicographically sort the properties by name
- double quotes used on all names
- double quotes used on all string values
- no space, or one-space, between names and the colon, and between the colon and the value
- no spaces between values and the following comma
- all other white space collapsed to either a single space or nothing – choose one
- exclude any properties you don’t want to sign (one example is, the property that holds the signature itself)
- sign the result, with your chosen algorithm
You may also want to think about how to pass that signature in the JSON object – possibly establish a well-known property name, like “nichols-hmac” or something, that gets the base64 encoded version of the hash. This property would have to be explicitly excluded by the hashing algorithm. Then, any receiver of the JSON would be able to check the hash.
The canonicalized representation does not need to be the representation you pass around in the application. It only needs to be easily produced given an arbitrary JSON object.