Kairouz et al. (2019) — ARXIV-1912.04977
Key takeaways
- Defines federated learning as a field and, most usefully, splits it in two: cross-device (very many unreliable clients with little data each, mostly unavailable, never individually addressable) versus cross-silo (few reliable organisational participants with a lot of data, always available, individually identifiable)
- Almost every practical concern differs between them — client sampling, communication cost, dropout handling, the threat model, whether a participant can be held to an agreement — and the survey contrasts both against traditional single-datacenter distributed learning
- Addresses cross-device by default, “though many of the problems apply to other FL settings”
- Describes the field as already deployed rather than prospective, and catalogues open problems across privacy, security, system design, fairness and efficiency
Standing in the anthology
Skimmed, not read — NOTE-074. A ~66,000-word multi-institution survey, and this reading covers its framing and taxonomy rather than its problem catalogue. Under ADR-025 that is not enough to source a practice from.
The third use of Skimmed in this record, and the first that is a judgement
about proportion rather than availability: reading a subfield survey in full,
to correct four generic bullets on a document nothing cites, is not a good use of
a reading pass. The taxonomy is the part with reach and it is now recorded.
What it reaches: the record’s distributed practices (SOTA-017, SOTA-018,
SOTA-019, and the parallelism in LIT-065 and LIT-102) all quietly assume a
datacenter — reliable, addressable, homogeneous workers on a known interconnect.
SOTA-155 is the one practice that does not, and the cross-device/cross-silo
distinction is the vocabulary for saying which assumptions a distributed practice
depends on. The record has no such vocabulary.
This document’s standing should be decided rather than read further. It is
Active in an anthology that does not track its field, nothing cites it, and the
one idea worth having from it is now in the note.
Audited by #121; unsourced because it has not been read.
Skimmed, no recommendations extracted. A 2019 survey of an area the record
otherwise does not cover, which is a scope question rather than a filing one.