As I understand it, a fundamental of Chaum's mix-net is that, absent an external adversary who can analyse traffic on links within the network, no mix can link the source and destination of any message unless it collaborates with every other mix participating in the cascade through which that message passes.
Most of the mix-net projects about which I have read (both those that have been implemented and those that remain theoretical) mitigate such a "collaboration attack" by introducing randomness into path selection: thus an adversary must collaborate with a significant proportion of all mixes throughout the network in order to achieve a high probability that those in any given cascade will all be collaborators. Various techniques have been suggested to further frustrate an adversary who attempts to so bias the network, such as viewing all mixes on a single subnet as potential collaborators.
If one is certain (at least, as much as one can ever be) that some particular mixes are honest (i.e. not collaborating with an adversary), it would at first glance appear desirable to ensure that at least one such honest mix is included in every cascade used for one's messages so as to significantly reduce (if not completely eliminate) the threat of such "collaboration attacks".
However, forcing all one's messages to pass through a subset of the network that is defined by one's trust relationships leaks information about those relationships. Over time, an adversary might be able to use such information to link together all one's messages or even identify oneself.
My question is: how can one measure this trade-off? Is there existing research that already has considered it (I have not been able to find any yet)?
Where senders are themselves also acting as mixes in the presence of sufficient inbound cover traffic, does the inclusion of known honest mixes in downstream cascades for their own messages actually provide any additional anonymity given that those messages could arguably have been forwarded for another party?