New Privacy Method Enhances Collaborative Learning Security

Collaborative learning often faces a dilemma between data utility and privacy, especially when multiple participants contribute data. Geometric Data Perturbation (GDP) offers a one-shot method where participants transform their private data into distance-preserving representations before uploading them. However, a common technique called shared-anchor alignment, used to make these representations compatible and improve model performance, introduces a critical vulnerability: if an analyst colludes with some participants, they can precisely reconstruct the private data of non-colluding parties. Previous attempts to mitigate this by adding noise directly to private data representations significantly degraded the utility of the learned models. This research proposes a new strategy: adding noise specifically to the anchor representations. Each participant independently transforms their private data and a shared anchor matrix, then perturbs only the anchor representation before uploading both. The central analyst then uses these noisy anchor representations to align the private-data representations via a Generalized Orthogonal Procrustes Problem. Experiments on datasets like MNIST and CelebA demonstrate that this "noisy-anchor alignment" achieves higher learning accuracy than private-data noise, while maintaining comparable privacy leakage under specified collusion models, thus offering a more favorable balance between privacy and utility.