I propose something that has been discussed many times on the forums - implementing connection encryption to the bitcoin protocol.

Details:
- Like SSH, people will be able to connect to each other securely, using cryptographic keys / identities they need to exchange first.
- As in SSH, each host will generate its fingerprint, and set of private/public keys. So nodes connecting will be able to verify each other
- Some nodes (like banks) will be able to publish their keys/fingerprints somewhere (like on their site), so that everybody can confirm who they are when connecting to their bitcoin clients
- When key/fingerprint of remote node changes, user will be warned & asked if he wants to connect anyway (like in SSH).

- To make things easier, one could use openssl library present in every major operating system to implement this.

Possible benefits:
- Possibility of having almost 100% anonymity for each node. (if some random traffic generator will be also implemented) A third party no longer will be able to tell who is who and which transaction is which by sniffing the traffic coming in/out of a node.
- Real security & more anonymity when connecting through TOR (right now the exit nodes can easily sniff/intercept all traffic, so using bitcoin on TOR is somewhat dangerous).
- Eleminate man-in-the-middle attacks.
- People will be able to create "semi-certificate authorities", which will store each node's fingerprints and show them publicly, so nodes of the network can be verified.
- Clients (like banks) could choose to connect only to trusted, SSL-verified nodes, and ignore the rest.

Possible disadvantages:
- Possibility of centralization of the network ?

Simple. So that nobody knows what IP were these transactions created by.

And SSH-like connection will be almost exactly the same as a SSL-like connection, but without central certificate authorities. Of course some mechanism of reading fingerprints from a HTTPS site could also be implemented, so people could create "semi-certificate authorities" which would be simply a list of node IP's with their fingerprints.

That wouldn't change anything.

Encryption is not pointless, because it is unlikely an attacker can control enough bitcoin nodes today.

Furthermore, it is nice to not be observed when I am submitting a new transaction to the network.  Those in the coffee shop have no business knowing that I am submitting a new transaction, even if the TX is propagated in the clear throughout the network.

Actually the reasoning behind not needing the encryption is because we always assume an adversarial peer we're communicating through. The only thing that should be considered is reducing timing attacks which would enable to track back the origin of a client, but that's a pretty slim chance, and partially solved by using Tor.