There isn't much complexity and performance cost of optimistically using a faster hash, and dynamically falling back to siphash (or another secure keyed hash fu

There isn't much complexity and performance cost of optimistically using a faster hash, and dynamically falling back to siphash (or another secure keyed hash function). Using associative arrays implemented via chaining, one could have a bit in the associative array header that indicates if the associative array uses the fast keyed hash or siphash (or another secure hash). When inserting an item, if one finds that the chain length exceeds a certain limit without the load exceeding the resize threshold (indicating poor hash distribution), one could rehash all of the keys using siphash. An associative array using open addressing could do something similar, switching hash functions if the probe sequence got too long (analogous to a chain being too long). Alternatively, if using open addressing, one could use something like cuckoo hashing. A very fast keyed hash could be computed, and upon a collision, siphash could be used, with a probe sequence similar to used by Python dicts. Though, if your use case has lots of missed lookups, then performance will be better just using siphash. Of course, one could use counters to detect this condition, set a bit in the associative array header to indicate all future lookups should just use siphash, and rehash all of the existing keys.