That's very interesting to hear. How about this?
The neural network takes decision based on weights that is given to each and every input. This is what i was thinking when i posted this. I'll just write it down in steps, Please let me know if i'm correct.
 
Step 1 : So at the beginning we set the same amount of weight to all the inputs( Taking good nodes and bad nodes as the inputs).

Step 2 : We make this model the administrator of the network which has equal amounts of good nodes and the bad nodes. and there is a pool of other nodes(good) which could be introduced into the network based on the actions in the network.

Step 3 : Now, We just let the network run. The algorithm is written in such a way that once a bad node tries to manipulate some node he's immediately replaced with a good node from the pool.

Step 4 : We could also give weights to attacks. If it's irrelevant and if it's making no much damage let's just keep the node. If it again repeats the same thing the node would be replaced.

Step 5 : For scalability we could just choose a particular set of nodes and simulate a tremendous amount of transactions and pin point to what kind of nodes are causing this and we could drastically reduce or try applying compression algorithms to that particular node.

I don't know if what i'm saying makes sense. This is what i've been trying to build from the past week