What prevents someone from using a Precomputed POW Spam Attack against the network?
A theoretical attack could use pre-made transactions to be released before their network expiry dates.
Rounds of fictitious transactions can be created and post-dated two weeks to three weeks in the future within an internal database and tagged by expiry date. These can be formed into subtangles if desired.
For example, a person makes 2.5 million to 5 million pre-made transactions that can be harvested at the right time, before each transaction will no longer be accepted by the network.
A simple network graph is constructed and 5 or 6 IRI nodes are created to be part of the network as regular Field/Nelson nodes but with a few software tweaks to ensure that they also listen to a private master node and relay those transactions first.
Within the second generation 45 neighbors will be broadcasting. By the time it reaches the third generation, 320 neighbors will be gossiping. By the time it reaches the fourth or fifth generation, the entire network will be permeated.
The gossip makes it impractical to track the origin of bad transactions - the network has effectively self-obfuscated its own security backups.
I think that pretty much sums up the basics of what is needed for this to succeed.