Just to follow-up to my earlier post....

1) There is a possibility for mining in a pool, and withholding blocks, to be net profitable for an attacker solely by the effect on the difficulty level, and delaying the difficulty retargeting.  If one had a substantial hashrate (say 5% of the network) then by mining in a pool and withholding valid blocks, blocks will be found 5% slower, and the difficulty retargeting will be delayed by 5%.  So they could potentially gain a few hours of time submitting shares to the pool at a higher value per share due to the lower difficulty.  Also, the new after-retarget difficulty will be lower than if they were submitting blocks properly, so again the attacker will receive a higher value per share submitted to the pool after the difficulty retarget.  There would be no reason for such an attacker not to use 100% of his hashrate in this way; but (for detection avoidance purposes) it would of course be smart to split it up among different pools, and different accounts within those pools.  The net effect will be the same.

2) There is a race condition that occurs whenever two miners find blocks at roughly the same time (i.e., a short fork in the blockchain, resulting in an orphaned block).  These events are rare; however, they do occur with predictable frequency.  This is the only time when the results of one miner can impact the results of another.  However, in the absence of persistent advantages in network capability, it is hard to imagine that these effects will be biased in one way or the other.  If there is a net effect, it will be much more subtle than the result presented in the paper.  Mining is still a series of independent trials, for all practical purposes.

3) I PM'ed ncourtois to see what he has to say.  I noticed that he has updated the paper with a new version, which discusses the block withholding attack on Eligius and asserts that the attackers "profited" from the attack (i.e., had greater earnings than if they had been submitting blocks properly).