For example, you can use outside information to run overly specific queries: “list the male students in woman’s studies 326 that got over a C”. Tricking me into running this, and observing the number of bytes returned will give you a nice approximation of Jim’s grade if you know he’s the only male in that class. The set of queries I can safely provide varies with the side channel information the attacker has.
Also, if you have the ability to subtract and compare (I think those are the right two operators), then you can use that to decrypt the underlying data.
There is also the best picture I have ever seen on wikipedia: https://en.m.wikipedia.org/wiki/Block_cipher_mode_of_operati... Scroll down to “electronic code book” to see why basic equality computationa break encryption guarantees.
To make things worse, the attacker can keep track of what time each encrypted block is read/written, and infer communication patterns between users, and even infer which blocks contain information on which topics. (This is similar to the concerns over governments that gather “metadata”, because gathering “data” would be illegal)
Anyway, I skimmed this article, but have spent days reading negative research results in this space that are pitched as though they are somehow practical, and I suspect I’m not the only one suffering from homomorphic encryption fatigue.