Novelty check: nobody is evaluating this¶

# research(2026-09). Searched for a benchmark or evaluation that asks whether the part a federated adapter scheme transmits identifies the client or organization that produced it. Nothing matches.

What exists nearby, and why none of it covers the question:

• The FlowerTune LLM Leaderboard is described as a first-of-its-kind public cross-domain benchmark suite for federated fine-tuning of LLMs. It scores utility across domains. A scheme that leaks its participants perfectly can top it.
• Membership-inference work in federated learning (FedMIA, CVPR 2025; the ACM Computing Surveys treatment, 10.1145/3704633) asks whether a data point was in training. The question here is whose update this is, which is a different target with a different exchangeable unit.
• The personalization literature measures what its split does for accuracy and communication cost, and states privacy as an architectural property rather than measuring it.

So the gap is not that the field lacks another partition scheme. It is that a deployment choosing among FedSA-LoRA, PFAdapter, SDFLoRA, FDLoRA, FedDPA, SecureGate and FedAMoLE has no way to ask which of them resists source attribution, and the three measured so far do not.

A correction to the table above. FedAMoLE is published at The Web Conference 2026 (doi 10.1145/3774904.3792147), not an unrefereed preprint as the arXiv identifier alone suggested. Its data-dependent expert assignment is therefore a refereed design, which strengthens rather than weakens the point that a structure chosen from a client's data distribution is a channel no weight-space partition covers.