chapter 12: xor-ing hash codes is discouraged and a potential security flaw

[adamwadesmith]

I understand using xor for combining hashes is useful pedologically, but it has problems and shouldn't be used in production code.

Problems:

1. XOR is commutative and associative, so it is a poor choice for ordered aggregates (list, tuples, strings, etc)
2. Duplicate cancelation is severe, since h(x) ^ h(x) = 0. In your vector example, all diagonal vectors will hash to 0.
3. Security: attackers can create collections with identical combined hashes, which would make all dictionary lookups become linear scans.
4. Small changes should flip many output bits in a way that looks random.

One solution is a Boost lineage custom combiner, though it also isn't resistant to attacks:

class MyObj:
    def __init__(self, a, b, c):
        self.a, self.b, self.c = a, b, c

    def __hash__(self):
        seed = 0
        for v in (self.a, self.b, self.c):
            h = hash(v)
            seed ^= h + 0x9E3779B97F4A7C15 + ((seed << 6) & ((1 << 64) - 1)) + (seed >> 2)
            seed &= (1 << 64) - 1  # keep bounded like a machine word
        # Convert to Python int in signed-ish way (optional)
        if seed >= (1 << 63):
            seed -= (1 << 64)
        return seed

Or just put a warning on the code.

Adam Smith