The whitespace is in who is NOT filing. A PatentBear sweep for federated-learning and privacy-preserving training in the 2021 window returns hundreds of records, but the assignee facet tells a striking story: the largest bucket by far is unassigned — individual inventors and small entities — with the named-corporate share led by IBM, Meta, Google, and Intel at modest counts. For a technique this strategically important, the big-lab footprint in the granted/published record is thinner than you'd expect.
A concrete exemplar of the corporate side is IBM's US11494700B2, "Semantic learning in a federated learning system," granted in late 2022 on a 2021-era application, CPC G06N 20/00 with NLP code G06F 40/30. It represents the pattern: incumbents filing specific federated-learning system claims, while the broader space fills with smaller players.
Federated learning matters because it promises model training without centralizing raw data — each device or institution trains locally and shares only updates. That maps directly onto privacy regulation and onto sectors (health, finance) where data cannot leave its silo. The CPC clustering reflects this: G06N 20/00 and G06N 3/08 for the learning, plus privacy/crypto codes like H04L 9/008 and G06F 21/6245 for the protection layer.
The whitespace read, with caveats: application counts are not granted rights, and a heavy unassigned bucket can reflect academic and startup activity that later gets acquired or assigned. But as of the 2021 record, there is visible room in the corporate landscape — the foundational privacy-preserving-training claims are not yet locked up by a handful of hyperscalers the way transformer and accelerator IP increasingly is.
For the IP strategist, the actionable point: federated learning in 2021 looks less consolidated than other AI subfields. Anchor any whitespace thesis to specific records like US11494700B2 and remember the counts are filing-date sensitive — the landscape is still forming.