There is little benefit to negotiation over the Bitcoin network protocol for traditional CoinJoin's besides eliminating the need for an additional networking layer.

On the downside, adding additional messages to the network protocol is likely an irksome process, and is not very flexible. A separate network may be rapidly iterated upon, and other shared transactions other that traditional CoinJoins may be added.

In regards to Tor, for Java there exists the Orchid library, which allows Tor to be easily integrated within Java applications. The main benefit of using Tor Hidden Services (to me at least, if I am understanding things correctly) is not really anonymity, but rather NAT traversal. Without Tor, you have to keep a port open to allow users to connect to you node and perform a decentralized CoinJoin. Tor hidden services connect to Tor Relays, and therefore do not require any ports to be open. As long as the NAT/firewall allows outgoing Tor connections, everything works out.

EDIT:
I forgot to mention, a downside of using Tor is that TomP2P and all other Java DHT libraries that I know of require ports to be open to ensure the integrity of DHT (if no nodes are hosting the DHT information, what's the point?). As such, in order to make the DHT robust the code would have to be extended to facilitate Tor Hidden services. This doesn't even address the fact that using a DHT to facilitate CoinJoining between number of users n>2 is a real pain.

Hence, decentralizing peer discovery is a job for another day week month.

Currently peer discovery is implemented with a centralized server. The server waits for N users to connect, then sends a message containing the IP Address and port of all participants. This approach is vulnerable to denial of service and is a single point of failure, but on the up-side any compliant server can be used. I still believe distributed peer discovery is ideal, but that can always be added later.

The centralized method is also NAT-friendly if Tor is used. Here is an idea for anonymous peer discovery and communication:

1. Each participant starts a Tor Hidden Service.
2. Using Tor, each participant connects to a peer discovery server, which is itself a Hidden Service. It announces the ID of its Hidden Service and open port.
3. The server then sends each participant a list of the Hidden Services. The participants then connect to these Servers and proceed with the decentralized CoinJoin process.

+ No traffic ever leaves the Tor network
+ No port forwarding / NAT traversal is required (in this sense it is more user-friendly than a non-anonymous

It should be noted that in order to prevent inputs and outputs from being linked by participants more complicated measures such as the blind signatures discussed on the first page must be used.

P.S.
Here is an example of a 10-way CoinJoin I generated using my library:
http://tbtc.blockr.io/tx/info/894d10fea8e017789e80e2965d3421572e42e19ba8c6f51ce4a22b3c40b0f831

This is similar to what a CoinJoin transaction would look like in practice, except a more secure implementation would mix the outputs around better.

I am currently working on a Java library which facilitates decentralized CoinJoin-ing using a BitcoinJ backend.

As of this moment the library only works with fixed CoinJoin participants, I have yet to implement participant discovery (that's next on the list).
It also is very insecure at the moment, as I hacked together some sections in order to test general principles. It will be some time before the code is solid enough to not be embarrassing :\

Current plan:
1. Implement peer discovery (perhaps using a DHT)
2. Fix error handling and enhance verification of transaction components
3. simplify, refactor, rename
4. post source code (under a permissive OSS license)
5. Make CoinJoin process more anonymous.
6. ALPHA release?

example:
http://tbtc.blockr.io/tx/info/c4d86d7a054e5979172b223a15d5d9594f703d6376ab294ee4b2da45ff77b0eb

This is a test CoinJoin transaction between only 2 users. In this example I set the change and output address to be the same. The general caveats of a CoinJoin transaction still apply: each change address is clearly linked with an output address, and therefore by using blockchain analysis it may still be possible to link addresses. True anonymity requires minimal address reuse and tools for managing taint.

For now, though, I just send an unconfirmed transaction to a new address of fixed size, then use that unconfirmed transaction as part of the CoinJoin. Needless to say, for this scheme 0 confirmation coinjoins should not be accepted! Regardless, I am trying to write the library to be adaptable as possible to different types of CoinJoins, including coinjoins where each user has multiple inputs, casual coinjoins, and coinjoins without any change address.

I have an idea about a way to perform a decentralized Coinjoin so that individual participants are unable to map inputs to outputs - without the need for an anonymity network or blinded signatures.

As a disclaimer, although I have experience programming I am not well-practiced in cryptography, so forgive me if I make any egregious mistakes and waste everyone’s time.

Given the IP addresses and public keys for all N Coinjoin participant nodes, I envision the following onion routing protocol (inspired by what is used in Tor):

1.   Decide upon a random path which visits every node once and ends up back at our node.

2.   Using Onion Routing, send a multi-level encrypted message along this path. Each node by using its private key to decrypt the message will be able to see where the message was supposed to originate from, the bitcoin addresses to be used as inputs, and where to send the message to next. The rest of the message will only be able to be further decrypted by the next destination. When the message gets back to us, it should have visited every node in the path.
See: http://en.wikipedia.org/wiki/Onion_routing

3.   Same protocol, but with output addresses (can be done in parallel with the input addresses).

4.   Same protocol, but with our signature of the completed transaction.

5.   Broadcast transaction.


In theory, each node will be unable to tell from what node the input and output addresses originated. However, I see several serious issues with this proposal, and would welcome even more critique:

• Timing Attacks: If a node receives the message early in the process, it is more likely that that the sender owns the associated addresses. This may be mitigated through random delays or other, more clever schemes.
  
• Message Size: A node may be able to analyze the size of the message to determine how far along it is. To counter this, a randomly-sized allotment of junk data should be included in the inner-most message.
  
• Slow: this protocol takes N time, as the message must be forwarded to each node.
  
• DDOS: What is to stop some other node from screwing everyone over? It is possible to see if our message was tampered with (using random numbers, accumulating counters, etc.), but it would still be difficult to make the protocol resistant to malicious nodes mucking everything up and wasting everyone’s time.
  
• Sybil Attacks: You’re pretty much screwed. Tor has similar issues, this is one thing you really can’t do anything about (other than favoring IP addresses that are likely to be in physically separate locations)
  
• Other stuff: It’s probably out there, I just can't think of it at the moment.
  

Well, what about the positives? Assuming all the negative obstacles are surmountable:

• Inputs are not linked to IP addresses.
  
• Outputs are not linked to IP addresses.
  
• Inputs are not linked to outputs any more than can be determined from looking at the finalized transaction.
  
• No anonymity network required. This is important as P2P over, say, Tor is a pain.
  
• No reconnecting required. You are not required to meet up with the same nodes again.
  
• If you are behind a NAT router or firewall blocking inbound connections, you can hire an open node as a proxy. As only you have access to your private key, you can have this node forward/receive your messages without worrying about it snooping (granted, it can still cause as much trouble as the other nodes, and you probably have to compensate the proxy owner by including a small transfer in your transaction).