Normally concept or dataset drift is unquantifiable in practice. The only ways to calculate it are with optimal transport techniques that take O(n^4). The proxies some ML practitioners use are unreliable when applied to security. This talk presents a way to calculate a concept drift number that takes O(log n). It is faster than most inference, so can be put inline in an ML pipeline. Also, as the structure used to calculate the drift is so cheap we can apply it per-user as an extremely effective defense against attacks.