Malicious server, apart from possibility of removing data, would look pretty unusual for the client (like returning incorrect data etc). So it's a good idea to

Malicious server, apart from possibility of removing data, would look pretty unusual for the client (like returning incorrect data etc). So it's a good idea to look for these patterns. An attack would probably look like trying to use the client as an oracle, and that's pretty detectable. But as MacLane pointed out in another comment this is only the case for the standalone database described in the ArXiv paper. The Hadoop scheme works quite differently