AI History Battle

fairness

The statistic that can't betray you

It is 2006 at Microsoft Research, and the comfortable fictions of "anonymization" are collapsing one re-identification at a time: stripped names undone by zip code and birthdate, anonymous search logs unmasked by the queries themselves. The lesson is that ad hoc scrubbing cannot be patched — privacy needs a definition with a proof attached. Publish aggregate statistics from a medical database with a mathematical guarantee that no individual's presence or absence can be inferred from the outputs, no matter what auxiliary information an adversary holds — and price the accuracy cost explicitly, because the guarantee is purchased with calibrated noise and the budget is finite. National statistics agencies and every hospital data-sharing agreement are waiting on this calculus. Anything weaker is a promise, and promises have been breaking.

privacy calculusprove

Who this problem belongs to

The two figures whose methods fit it best, out of 30 in contention.

b. 1958 · ai-classic
99

This is Dwork's field, essentially by definition. Working at Microsoft Research in the mid-2000s, she co-invented differential privacy specifically to replace ad hoc anonymization with a mathematical definition carrying a provable guarantee: no individual's presence or absence in a dataset can be inferred beyond a quantified bound, regardless of an adversary's auxiliary knowledge, purchased with calibrated noise against an explicit privacy budget. That is a description of this problem's ask, not an analogy to it. She built the composition theorems, the noise mechanisms, and the formal privacy-accuracy tradeoff the problem demands be priced explicitly. No carrier's toolkit is closer to zero translation distance from this scenario.

b. 1975 · deep-modern
85

Song's work on AI security and adversarial privacy attacks (2000s-2010s, Berkeley) is squarely built around exactly the failure mode this problem opens with: re-identification attacks that undo naive anonymization by exploiting auxiliary information. She has studied both sides — how attackers unmask stripped datasets and how to build systems resistant to such attacks — giving her direct expertise in the threat model differential privacy formalizes against. Her focus skews toward adversarial and systems security more than the specific mathematical mechanism-design (noise calibration, composition) that produces the provable guarantee itself, keeping her just behind Dwork in this specific mechanism-design task.

In the mind map

The same ideas, as concepts rather than history — in John's ML knowledge map.

Differential Privacy

30 figures are scored on this problem. Draw it in a battle to see where you land.