The problem I think is that at some point the register containing the encrypted address needs to be written to the stack, where it could be overwritten. So if i

The problem I think is that at some point the register containing the encrypted address needs to be written to the stack, where it could be overwritten. So if it was just hashed, an attacker could overwrite both the hash and the plaintext return address, and return wherever they want. Although hashing with a secret salt would solve that. I don't seen any benefit over encrypting though. One attack I can see is in their model of "the attacker can read all memory" the attacker reads the memory at a time when system() is legitimately called, then the attacker can know the encrypted version of the system() address, and use that to overwrite some other return address with system() and the encrypted version. But I think it would be very unlikely an attacker would be able to do that.