How are you not getting this?
endpoint0-n ---> 'server' ---> endpoint0-n
where ---> is a random level of encrypted noisy signal. the endpoint does not open a connection to the other endpoint, it opens it to the 'server'.
no amount of traffic analysis or packet injection is going to mess with that.
the absolute best traffic analysis can do is provide a 1/n probability based on active connections to the 'server' that endpointA was talking to endpointB.
> we have to make the latency worse.
> there's a notion that latency is good for anonymity
we have to do something because of a notion?
i give up.
Pond does have that property, if you use the defaults. I don't think a lot of other systems do. The tradeoffs are pretty steep in terms of delaying interactions until the next time window, sending and receiving cover traffic in every time window, and accepting hard limits on your data rate that are bounded by the cover traffic.
What I think you can't do safely, for example, is say "I have a cover traffic pattern that is a Gaussian distribution of amount of data transmitted and received, and now I have to send a bunch of data, so I'll just pick the high of the distribution and send a whole bunch of data at once". One reason this is unsafe is that you'll simultaneously skew the recipient's distribution, creating a statistical signal that you and the recipient were communicating.
the traffic from the 'server' is completely random and unrelated to the traffic into it.
within this level of encrypted traffic noise, data is carried, smaller than the carrying capacity of that traffic.
The traffic analysis comes in where you notice correlations between increased or decreased activity on one link and a corresponding change in activity on another link. Just having noise isn't necessarily enough to spoil those correlations; after all, so much of modern statistics is about detecting very weak signals given many noisy observations.
There is also research about active attackers shaping traffic flows (like delaying or blocking packets injecting additional packets). Then the notion is that the changed shape of a flow will be visible elsewhere on the network, and that's the destination. Unfortunately, this seems to work really well!