ε-Almost Δ-Universal hashing #
ε-almost-Δ-universal₂ (ε-A∆U₂), due to [BKST15],
Definition 1.1, generalises ε-AU₂ by
bounding all difference probabilities, not just the zero-difference (collision) probability.
It requires [AddCommGroup Output] to express H s x − H s y.
Special cases of this definition are known by several names in the literature.
In particular, if G is the group of length-l strings of bits under element-wise XOR operation,
this case has been called
-"ε-otp secure" (e.g. Krawczyk94, 10.5555/1755009.1755041)
- ε-almost XOR universal (e.g. Rogaway99, Rogaway1999)
The chain of implications is:
almostStronglyUniversal2 ε → almostDeltaUniversal2 ε → almostUniversal2 ε
The left implication sums the joint bound over all a : Output to bound each difference.
The right implication specialises b = 0 (since H s x − H s y = 0 ↔ H s x = H s y).
HashFamily.almostDeltaUniversal2 ε H(defined inUniversalHashing.Basic): H is ε-A∆U₂ if for all distinctx ≠ yand everyb : Output:Pr_{s}[H s x − H s y = b] ≤ ε. [BKST15], Definition 1.1HashFamily.almostDeltaUniversal2_mono: ε-A∆U₂ is monotone inε.HashFamily.almostUniversal2_of_almostDeltaUniversal2: A∆U₂ implies AU₂.HashFamily.almostDeltaUniversal2_of_almostStronglyUniversal2: ASU₂ implies A∆U₂.
A perfectly Δ-universal family is (1 / |Output|)-A∆U₂ — with the optimal parameter,
since for fixed x ≠ y the probabilities Pr[H s x − H s y = b] sum to 1 over b.
ε-A∆U₂ is monotone in ε: if H is ε₁-A∆U₂ and ε₁ ≤ ε₂, then H is ε₂-A∆U₂.
almostDeltaUniversal2 ε implies almostUniversal2 ε: specialise b = 0, since
H s x − H s y = 0 ↔ H s x = H s y.
[BKST15], Definition 1.1, remark
almostStronglyUniversal2 ε implies almostDeltaUniversal2 ε:
summing the joint bound Pr[H s x = a ∧ H s y = a − b] ≤ ε / |Output| over all a
gives Pr[H s x − H s y = b] ≤ ε.
This makes the full implication chain
almostStronglyUniversal2 ε → almostDeltaUniversal2 ε → almostUniversal2 ε explicit.
[BKST15], Definition 1.1; implicit in the paper's discussion
ε-A∆U₂ implies (ε · |Output|)-ASU₂: each joint probability
Pr[H s x = a ∧ H s y = b] equals Pr[H s x − H s y = a − b]
(same event after fixing one marginal), which is at most ε by A∆U₂.
The ASU₂ bound (ε · |Output|) / |Output| = ε then follows.