The field states source-hiding as a property and still does not measure it¶

# research(2026-09). A sweep of this year's federated-LoRA privacy work, read for whether anyone has now measured what RQ2 measures.

• Fed-DiffLoRA (IEEE Trans. Image Processing 2026) splits a client's adapter into orthogonal content and style subspaces and fuses the style half "while suppressing client-identifiable information". It is the first design in this family to name source-hiding as a goal, and it offers no attack against which the suppression is shown. Text-to-image, so the apparatus here cannot run it; it is the citation for why the measurement is missing.
• AS-LoRA (arXiv:2605.05769) picks A or B per layer and per round from a curvature score. Since A alone is the best identifier in our factor readings, a cut that moves between rounds is the case where leakage is not a property of the design at all.
• Rethinking LoRA for Privacy-Preserving Federated Learning (ICLR 2026) is about utility under differential privacy: gradient coupling, noise amplification, sharpness. No identifiability measurement, which is the pattern.
• LoRA as Oracle (arXiv:2601.11207) reads as source attribution in search summaries and is not: it audits a received model for backdoors from the geometry of a low-rank update. Recorded because the summary is misleading and the title will come up again.

Nothing measures whether a transmitted adapter identifies its source at an operating point an attacker would use. The claim RQ2 registers is still unoccupied.