The place where it really starts to make sense to think about Noise is when you're already committed to doing a custom protocol. Jason Donenfeld's Wireguard is a great example: a fresh protocol makes a lot of sense for a new VPN transport.
Ironically, in the context of a secure messenger, the Noise/TLS distinction isn't really important at all. The important cryptography in an E2E-secure messenger happens at a layer higher than the transport itself.
Later
Spitballed this question on crypto dork Slack and there is a use case with a clear answer: if you're doing a datagram transport rather than TCP streams, Noise is probably always a clear win. DTLS is a mess.
I was also told that Wireguard uses Noise on top of UDP, so it would be interesting to see what they did there.
Another good point for Noise: if you're planning to implement a TLS-like protocol from scratch, I would go for Noise instead of TLS. It's just simpler, lighter and easy to customize to your own needs.
But indeed DTLS is a mess for other reasons.
For WireGuard I use Noise with stronger DoS resistance and anti replay measures than DTLS, custom tailored for being a layer 3 tunnel. Details here: https://www.wireguard.io/papers/wireguard.pdf
In what way is MITM easier in UDP than TCP? Assuming you use similar techniques to verify your connection, I can't see a difference.
- The interactions in the protocol are precisely specified
- The exact security properties of the different interactions are precisely specified, including in-depth concepts like the AKE's KCI properties (AKE: authenticated key exchange, KCI: key compromise impersonation, where post-compromise, an attacker can impersonate anyone to the victim, instead of being able to just impersonate the victim to anyone).
- Several mutually compatible reference implementations.
As great as it is, most applications should still rely on TLS for transport security. Noise is primarily for places where TLS' properties aren't suitable.
Another interesting tidbit that might not be obvious: like TLS, Noise is one protocol with many instantiations. Because it builds on a few simple primitives, these can be swapped out, although Noise defines sane defaults. However, an application built on top of Noise is far more likely to have one fixed ciphersuite set (agility, but not negotiation) -- so it wouldn't be unreasonable to think of Noise as a blueprint for a set of protocols.
I'd be happy to answer any detailed questions about the Noise protocol; I am a cryptographer, but I've had nothing to do with its development.