Patterns: authentication

The server generally authenticates to the client by its SSL certificate. One very popular way of doing this is to have the client trust a well known authority, a certificate authority (CA) and then verify that the name of the server’s certificate matches expectations, and that the server’s certificate is signed by the CA and not yet expired. This is an excellent method if previously unknown clients do drive-by connections to the server, but has no real value if all the clients are pre-approved for a particular service on a group of particular servers. In that case, it’s just as easy, much cheaper, and arguably more secure, to provide all the clients with the server’s public certificate ahead of time, during installation, and let them verify the server certificate against that public key. 

But how do we authenticate a client to the server? Well, we provide the server with the client’s public key during installation and then verify that we have the right client on the line during connections.

We could let the client authenticate against the server using the built-in mechanisms in HTTPS, but the disadvantages of that are numerous, not least of which is getting it to work, and maintaining it over any number of operating system updates. “Fun” is not the name of that game. Another major disadvantage is that the protection and authentication that comes from the HTTPS session are ephemeral; they don’t leave a permanent record. If you want to verify after the fact that a particular connection was encrypted and verified using HTTPS, all you have to go on are textual logs that say so. You can’t really prove it.

What I’m describing now is under the precondition that you’ve set up a key pair for the server, and a key pair for the client, before time and that both parties have the other party’s public key. (Exactly how to get that done securely is the subject of a later post.)

Step 1

The client creates a block consisting of:

  • A sequence number, which is one greater than the last sequence number the client has ever used.
  • Current date and time
  • Client identifier
  • A digital signature on the above three elements together, using the client’s private key

The client sends this block to the server.

Step 2

The server (front-end) then uses the client identifier to retrieve the client’s public key, verifies the signature, and checks that the sequence number has never been used before. The server also checks that the date and time is not too far in the past or the future. 

If everything checks out fine, the server records the session in the database and updates the high-water mark (last used sequence number from this client).

Step 3

The server creates a block consisting of:

  • The client’s sequence number
  • Current date and time
  • Client identifier
  • Server identifier
  • A digital signature on those elements together, using the server’s private key
This block is then sent to the client, allowing the client to verify the signature, and save the record to its own local database. Since the block contains the client’s sequence number, which needs to match, it cannot be a playback.

Benefits

Doing it this way creates a verifiable record in the database about the authentication. The signature is saved and can be verified again at any time. This allows secure non-repudiation.

Creating the authentication as a signed block also means that the client does not necessarily need to communicate directly with the server. If a client needs to deliver documents to another system, which in turns forwards them, it can also deliver the authentication block the same way. The forwarding system does not need to hold any secrets for the actual client to be able to do that. This allows us any number of intermediate message storages, even dynamically changing paths, with maintained authentication characteristics.

I should also note that doing the authentication this way decouples the mechanism from the medium. If you replace HTTPS connections by FTP, or even by media such as tape or floppy disks (remember those?), this system still works. You can’t say the same of certificate verification using HTTPS.

Patterns: sacrificial front-end

Over the years, I’ve borrowed and invented a number of design patterns for projects of all kinds. Most, if not all, were doubtlessly already invented and used, but mostly I didn’t know that then. Most of my uses of these patterns are at least 15 years ago, often 20, but I’m seeing more and more of them appear in modern frameworks and methodologies. So this is my way of saying, “I told you so”, which is vaguely satisfying. To me.

Forgive the names; I have a hard time coming up with suitable labels for them.

Blackboard architecture

The Blackboard architecture is well-known. Or should be, except it seems I always need to explain it when it comes up. It has a lot of great aspects and results in effective and extremely decoupled designs. I’ll most certainly come back to it several times.

A blackboard is a shared data source. Some processes write messages there, while other processes read them. The different processes never need to talk to each other directly or even know of each other’s existence.

From this flows a number of advantages, a few of which are:

  • Different cooperating services can be based on entirely different languages and platforms.
  • The interaction is usually one way (compare to Facebook’s Flux and React), greatly simplifying interactions.
  • If done right, the data structures are immutable, elimination contention problems.Receiving messages

Hop, skip, and jump

Let’s get to my first blackboard-based pattern, namely how to protect a front-end machine from compromise. In this design, the front-end machine is an Internet facing computer receiving medical documents from a number of clients around the net. The documents arrive individually encrypted using the server’s public key. The front-end machine is assumed to be hacked sooner or later and we don’t want such a hack to lead to the ability of the hacker to get at decrypted documents or other secrets.

So, what we do is we let the front-end machine take each received and encrypted message and store it in an SQL database located in its own network segment. The message ends up in a table that only holds encrypted messages, nothing else.

Another machine on that protected network segment picks up the encrypted messages from the database, decrypts them, and stores the decrypted messages in another table.

Frontend1

The “front-end” machine is exposed to the internet, so let’s assume it is completely compromised. In that case, the hacker has access to all the secrets that are kept on that machine and has root. This would allow the hacker to do anything on that machine that any of my programs are allowed to do. 

The first role of the front-end machine is to authenticate clients that connect. We can safely assume that the hacker won’t need to authenticate anymore. 

The second role of the front-end machine is to receive messages from the client. These messages are then sent on to the database through a firewall that only allows port 1433 to connect to the database. The login to the database for the front-end machine is kept on the front-end machine so the hacker can use that authentication. However, the only thing this user is permitted is access to a number of stored procedures tailored specifically to the needs of the front-end. Among these stored procedures, there are procedures to deliver messages to the database, but not much more. There is most definitely no right granted for direct table access. In other words, the hacker can deliver messages to the incoming message table, but nothing else.

Behind the firewall there is another machine that has no connection to anything except the database. That machine has access to a small number of stored procedures tailored for its use, among which are procedures to pick up new incoming messages and deliver decrypted messages back to the database.

These crypto servers first verify that the message it picks up from the database carries a valid digital signature from a registered user system, and only then does it decrypt the message with its private key. If the hacker on the front-end had delivered fake messages, these would be detected during signature verification and discarded.

With this design, the hacker on the front-end has just a very narrow channel through which to jump from the front-end to the database, namely through the port 1433 and the SQL server software itself. But let’s assume he succeeds, somehow. If we’re really paranoid, we’d split the database into two instances on different machines, completely isolated from each other and only bridged by the crypto servers as in the next image.

Frontend2

To get at the plain text content of the messages, the hacker in this case, if coming from the internet at least, needs to:

  1. compromise the front-end
  2. crawl through 1433 and compromise the database (or compromise the firewall, then the database server)
  3. via a tailored message, compromise the crypto routines on the crypto server
  4. get at the secure database

The crypto server does not have any communication to the internet whatsoever, so even if it ever got compromised, it could only be controlled through messages passing through the database, and would need to exfiltrate that same way. Not impossible, but hardly easy, either. The hacker would probably choose some other way to get at the secure database.

So, what about outgoing?

The outgoing messages follow the exact opposite path. Do I really have to draw you a picture?