What is the definition of security of universal hashing

As a cryptographic primitive, universal hashing should have somewhat a criterion on its security. How is its (computational) security defined? Or, in other words, what "breaks" a universal hashing?

There's a variety of formal definitions, let me illustrate by giving the three most common definitions. Let $\Pr_{h\in \mathcal{H}}[X]$ denote the probability of event $X$ if $h$ is chosen independently and uniformly at random from set $\mathcal{H}$. A hash function family $\mathcal{H}$ is $\epsilon$-almost universal ($\epsilon$-AU) if for any $m \neq m'$ we have

$$\Pr_{h\in\mathcal{H}}\left[h(m) = h(m
')\right] \leq \epsilon.$$

A hash function family $\mathcal{H}$ is $\epsilon$-almost delta universal ($\epsilon$-ADU) if for any $m \neq m'$ and any $\delta$ we have

$$\Pr_{h\in\mathcal{H}}\left[h(m) \ominus h(m') = \delta\right] \leq \epsilon,$$
where $\ominus$ is subtraction in some group (e.g. subtraction mod $n$ or XOR).

A hash function family $\mathcal{H}$ is $\epsilon$-almost strongly universal ($\epsilon$-ASU) if for any $m \neq m'$ and any $d, d'$ we have

$$\Pr_{h\in\mathcal{H}}\left[h(m) = d \wedge h(m') = d'\right] \leq \epsilon.$$

If $D$ is the set of hash digests, we find that $\epsilon/|D|$-ASU implies $\epsilon$-ADU which in turn implies $\epsilon$-AU.

Or, in other words, what "breaks" a universal hashing?

All three of the above three formal definitions area bound some probability. You break a universal hash by showing that the respective event $X$ happens more often than the bound claims.

StackExchange Computer Science Q#161613, answer score: 3