Who actually owns this subfield? Across the portfolio, the 2022 federated-learning record shows the corporate footprint thickening relative to 2021. A PatentBear sweep surfaces IBM, Meta Platforms, Google, and Intel as the recurring named assignees, layered over a still-large unassigned bucket. The structural finding: the systems-and-orchestration side of federated learning is consolidating toward incumbents, while the cryptographic privacy layer remains more dispersed.
A representative privacy-layer exemplar is US11429903B2, "Privacy-preserving asynchronous federated learning for vertical partitioned data," granted August 30, 2022 to JD's digits-technology arm, CPC G06N 20/20 with G06F 17/16. It illustrates the split: specialized privacy-and-partitioning claims held outside the big-four hyperscalers, in a more crowded field.
The subfield divides cleanly along its CPC seams. The learning-orchestration side clusters in G06N 20/00 and G06N 3/08 — how clients train and how updates aggregate — and that's where the incumbents are densest. The privacy side clusters in H04L 9/008 (homomorphic-style crypto) and G06F 21/6245 (privacy-preserving processing), and that's where ownership is more scattered across smaller and non-US entities.
The caveats are the usual two, and they bite here. Counts are filing-date sensitive, so a 2022 map reflects late-2010s-to-2020 investment. And application counts are not granted rights — a heavy presence in publications can evaporate during prosecution. So "who owns it" is really "who has filed and, in fewer cases, been granted," which are different questions.
For the strategist: if you're assessing freedom to operate in federated learning, treat the orchestration layer as increasingly incumbent-held and the privacy layer as more open, and anchor every ownership claim to specific records like US11429903B2 rather than to the aggregate facet counts.