Posts Tagged ‘ Opportunistic encryption ’

Basic blueprint for a link encryption protocol with modular authentication

The last few years we have seen more and more criticism build up against one of the most commonly used link encryption protocols on the internet, called SSL (Secure Socket Layer, or more precisely it’s current successor TLS, Transport Layer Security) for various reasons. A big part of it is the Certificate Authority issued certificates model of authenticating websites where national security agencies easily can get fake certificates issued, and another big part is the complexity who have lead to numerous implementation bugs such as OpenSSL’s Heartbleed and Apple’s Goto Fail and many more, due to the sheer mass of code where you end up not being able to ensure all of it is secure simply because the effort required would be far too great. Another (although relatively minor) problem is that SSL is quite focused on the server-client model, despite that there’s a whole lot of peer-to-peer software using it where that model don’t make sense, and more.

There’s been requests for something simpler which can be verified as secure, something with opportunistic encryption enabled by default (to thwart passive mass surveillance and increase the cost of spying on connections), something with a better authentication model, and with more modern authenticated encryption algorithms. I’m going to make a high-level description here of a link encryption protocol blueprint with modular authentication, that has been inspired by the low-level opportunistic encryption protocol TCPcrypt and the PGP Web of Trust based connection authentication software Monkeysphere (which currently only hooks into SSH). In essence it is about the separation and simplification of the encryption and the authentication. The basic idea is quite simple, but what it enables is a huge amount of flexibility and features.

The link encryption layer is quite simple. While the protocol don’t really have separate defined server / client roles, I’m going to describe how the connections work with that terminology for simplicity. This will be a very high-level description. Applying it to P2P models won’t be difficult. So here it goes (and to the professional cryptographers in case any would read this, please don’t hit me if something is wrong or flawed, please tell me how and why it is bad and suggest corrections so I can try to fix it);

The short summary: A key exchange is made, an encrypted link is established and a unique session authentication token is derived from the session key.

A little longer summary: The client initiates the connection by sending a connection request to the server where it initates a key exchange (assuming a 3-step key exchange will be used). The server responds by continuing the key exchange and replying with it’s list of supported ciphers and cipher modes (prioritization supported). Then the client finishes the key exchange and generates a session key and selects a cipher from the list (if there is an acceptable option on the list), and tells the server what it chose (this choice can be hidden from the network since the client can send the HMAC or an encrypted message or similar of it’s choice to the server). The server then confirms the encryption choice, and the rest of the encryption is then encrypted using that session key using the chosen cipher. A session authentication token is derived from the session key, such as through hashing the session key with a predefined constant, and is the same for both the client and the server, and the token is exposed to the authentication system to be used to authenticate the connection (for this reason it is important that it is globally unique, untamperable and unpredictable). Note that to prevent cipher downgrade attacks the cipher lists must also be authenticated, which could be done by verifying the hashes of the lists together with the session auth token – if the hashes is incorrect, somebody has tampered with the cipher lists and the connection is shut down.

And for the modular authentication mechanism:

The short summary: Authentication is made through both cryptographically verifying that the other end is who he claims to be and verifying that both ends have the same session auth token (it must not be possible to manipulate the key exchange to control the value of the session key and thus the session auth token). It is important that the proof of knowing the session auth token and the authentication is combined and inseparable and can’t be replayed in other sessions, so the token should be used as a verifiable input in the authentication mechanism.

A little longer summary: What type of authentication is required varies among types of applications. Since the authentication is modular, both ends has to tell the other what type of authentication it supports. A public server would often only care about authenticating itself to visitors and not care about authenticating the visitors themselves. A browser would usually only care about identifying the servers it connects to. Not all supported methods must be declared (for privacy/anonymity and because listing them all rarely is needed), some can be secondary and manually activated. The particular list of authentication methods used can also be selected by the application based on several rules, including based on what server the user is connecting to.

There could be authentication modules hooking into DNSSEC + DANE, Namecoin, Monkeysphere, good old SSL certificates, custom corporate authentication modules, Kerberos, PAKE/SRP and other password based auth, or purely unauthenticated opportunistic encryption, and much more. The browser could use only the custom corporate authentication module (remotely managed by the corporate IT department) against intranet servers while using certificate based authentication against servers on the internet, or a maybe a Google specific authentication module against Google servers, and so on. The potential is endless, and the applications is free to choose what modules to use and how. It would also be possible to use multiple authentication modules in both directions, which sometimes could be useful for multifactor authentication systems like using a TOTP token & smartcards & PAKE towards the server with DNSSEC + DANE & custom certificates towards the client. It could also be possible for the authentication modules on both ends to request the continous presence of a smartcard or HSM on both ends to keep the connection active, which could be useful for high-security applications where simply pulling the smartcard out of the reader would instantly kill the connection. When multiple authentication modules is used, one should be the “primary” module which in turn invokes the others (such as a dedicated multifactor auth module, in turn invoking the smartcard and TOTP token modules) to simplify the base protocol.

Practically, the authentication could be done like in these examples: For SRP/PAKE and HMAC and other algorithms based on a pre-shared key (PSK) both sides generate a hash of the shared password/key, the session auth token, the cipher lists and potentially of additional nonces (one from each party) as a form of additional challenge and reply-resistance. If both sides have the same data, then the mutual authentication will work. For OpenPGP based authentication like with Monkeysphere, a signature would be generated for the session auth token, both parties’ public keys and nonces from both parties, and then that signature would be sent stand-alone to the other party (because the other party already have the input data if he is the intended recipient), potentially encrypted with the public key of the other party. For unauthenticated opportunistic encryption, you would just compare the cipher lists together with the session auth token (maybe using simple HMAC together with challenge nonces) to make a downgrade attack expensive (it might be cheaper to manipulate the initial data packet with the cipher list for many connections so that the ciphertext later can be decrypted if one of the algorithms is weak, than to outright pull off a full active MITM on all connections).

I have also thought about how to try to authenticate semi-anonymously, i.e. such that neither party reveals who they are unless both parties know each other. The only way I think this is possuble is through the usage of Secure Multiparty Computation (MPC) and similar algorithms (SRP/PAKE is capable of something similar, but would need on average a total of x*y/2 comparisons of shared passwords if party A has x passwords and B has y passwords). Algorithms like MPC can be said to cryptographically mimic a trusted third party server. It could be used in this way: Both parties have a list of public keys of entities it would be willing to identify itself to, and a list of corresponding keypairs it would use to identify itself with. Using MPC, both parties would compare those lists without revealing their contents to the other party – and if they both are found to have a matching set of keypairs the other recognize and is willing to authenticate towards, the MPC algorithm tells both parties which keypairs matches. If there’s no match, it just tells them that instead. If you use this over an anonymizing network like Tor or I2P, you can then suddenly connect to arbitary services and be able to prove who you are to those you already know, while remaining anonymous towards everybody else.

It would even be possible for an application to recognize a server it is connecting to as a front-end for several services, and tell the authentication manager to authenticate towards those services separately over encrypted connections (possibly relayed by the front-end server) – in particular this allows for secure authentication towards a site that uses both outsourced cache services (like Akamai) and encryption accelerator hardware (which you no longer have to trust with sensitive private keys), making it cheaper to securely implement services like private video hosting. In this case the device performing the server-side authentication could even be a separate HSM, performing authentication towards clients on the behalf of the server.

The protocol is also aware of who initiated the connection, but otherwise have no defined server / client roles. Although the authentication modules are free to introduce their own roles if they want to, for example based on the knowledge of who initated the connection and/or who the two parties of the connection is. It is also aware of the choice of cipher, and can therefore choose to provide limited access to clients who connects using ciphers that are considered having low security, but still secure enough to be granted access to certain services (this would mainly be important for reasons such as backwards compatibility and/or performance on embedded devices).

The authentication module could also request rekeying on the link encryption layer, which both could be done using a new key exchange or through ratcheting like in the Axolotl protocol, or simply through hashing the current session key to generate a new one and deleting the old one from RAM (to limit the room for cryptanalysis, and to limit how much of the encrypted session data can be recovered if the server is breached and the current session keys is extracted).

But what if you already have a link encryption layer with opportunistic encryption or other mechanism that allow you to generate a secure session auth token? You shouldn’t have to stack another layer of encryption on top of it just to be compatible if the one you already are using is secure enough. There’s a reason the link encryption and authentication is separate here – rather than hardcoding them together, they would be combined using a standardized API. Basically, if you didn’t use the “default” link encryption protocol, you would be using custom “wrapper software” that would make the link encryption you are using look like the default one to the authentication manager and provide the same set of basic features. The authentication manager is meant to only rely on the session auth token being globally unique and secure (unpredictable) to be able to authenticate the connection, so if you can achieve that for the link encryption then you’re good to go.

(More updates coming later)

References:

http://web.monkeysphere.info/

http://tcpcrypt.org/

https://gist.github.com/Natanael90/556350

https://mailman.stanford.edu/pipermail/tcpcrypt-dev/2010-August/000007.html

https://whispersystems.org/blog/advanced-ratcheting/

http://www.metzdowd.com/pipermail/cryptography/2014-May/021475.html

Advertisements
%d bloggers like this: