A new arXiv preprint proposes a way to do collaborative learning without exposing participants' raw data. The method, called geometric data perturbation, lets each participant apply a secret transformation that preserves distances between data points, then upload the transformed representation. Because the transformation is distance-preserving, the geometry of the data remains useful for learning while the original values are obscured.
The paper also describes noisy-anchor alignment as part of the approach. While the abstract does not detail how this alignment works, it appears to address how separately perturbed representations can be made compatible for joint learning. The combination is framed as enabling one-shot representation sharing, meaning participants do not need to exchange multiple rounds of updates.
As a single preprint with a limited abstract, the source leaves several specifics unstated, such as the exact form of the transformation or experimental results. Still, the proposal points to a practical direction for privacy-preserving collaborative learning, where the trade-off between data utility and confidentiality is handled geometrically.