░░twurst.com / stun-without-trust (2018-03-08)

STUN Without Trusted Servers

I'm working on adding ENR support to go-ethereum. ENRs are self-signed documents that describe the capabilities and endpoints of nodes on the Ethereum peer-to-peer network.

To sign a document that includes the nodes IP address, we need to actually know that address. But computers behind NAT are unaware of their external, Internet-facing IP address. There are two common ways find this address:

  1. Asking the router using UPnP or NAT-PMP
  2. Using STUN with a known server on the Internet

The first option is simple but comes with a few disadvantages: the router might not speak UPnP or NAT-PMP, and it might know about the correct external IP either. This usually works on residential networks though.

The second option is great if you trust in a certain STUN service. However, we do not wish to trust in such a service in a peer-to-peer network setting. Instead, information about the IP should be provided by other nodes in the network. The rest of this article is a collection of ideas around replicating STUN functionality without the usual server in the middle.

Determining The Mapped Address

Our UDP-based protocol includes a field that echoes the observed IP and port back to the sender. Every time a request is sent, the corresponding reply states what the UDP envelope address of the request was. We can record those IPs as 'statements' of the form "node N says my IP is A at time T".

statements = {
    "node_1": ("203.0.113.2", 1519984067),
    "node_2": ("198.51.100.55", 1519984023),
    ...
}

We can compute a prediction about the current external IP from those statements. We can't put too much trust into any particular statement because its N might be acting maliciously. If the external IP is mispredicted and then published in a signed document we might loose contact with the network because nobody finds us with that IP.

Let's use a simple algorithm: The current prediction is the IP stated by the majority of nodes within a time window. If there is no clear majority, the predicted IP is undefined.

MIN_STATEMENTS = 10

def predict_ip(statements):
    count, maxcount, predicted_ip = {}, 0, None
    for (host, s) in statements.items():
        ip = s[0]
        count[ip] = count.get(ip, 0) + 1
        if count[ip] > maxcount and count[ip] >= MIN_STATEMENTS:
            maxcount, predicted_ip = c, ip
    return predicted_ip

Determining NAT Behavior

Not all NAT setups are alike. For our purposes, the only distiction we need is whether the NAT is a Full Cone NAT or not. A NAT with this property will relay all packets sent to the mapped port, even from hosts that we haven't talked to.

Sorry, your browser does not support SVG.
Figure 1: Illustration of Full-Cone NAT, from Wikipedia

With STUN, checking for this property can be done by sending a request to one endpoint while requesting that the server's reply should be sent from a different endpoint. If the reply arrives, the NAT can be considered Full Cone.

We can achieve this in by recording node identies which we have sent our record to:

contacted = {}

def on_send(host):
    contacted[host] = time.monotonic()

If a node we haven't recently contacted sends a packet to us, we know the NAT is capable of full cone translation.

def is_full_cone_nat():
    return any(host not in contacted for host in statements)