Progressive-Resolution Secure Aggregation for Federated Learning

Progressive-Resolution Secure Aggregation (PSA), introduced by Seyed Mohammad Azimi-Abarghouyi in arXiv:2610.00695v1 (September 30, 2026), is a novel protocol that allows a server to recover successively finer resolutions of a federated aggregate without requiring renewed client participation after the initial upload.

• Core Mechanism: PSA uses compatible nested-lattice representations to decompose quantized updates into bounded refinement layers, protected by separate secure aggregation and aggregate release pads that remain hidden from the server until authorized by a non-colluding controller.
• Key Advantage: Unlike conventional secure aggregation which fixes aggregate precision at upload, PSA enables a minimum-necessary-resolution policy, releasing coarse aggregates first and refining them only when justified, thereby reducing privacy leakage and communication overhead.
• Technical Cost: The protocol incurs an arithmetic-payload premium of approximately $(1/2)\log_{\rho}K$ bits per model dimension for each additional release layer (where $K$ is the number of clients), which is proven to be necessary within its specific interface class to handle overflow risks in modular sums.
• Performance: Federated experiments on MNIST and CIFAR-10 demonstrate that PSA matches the accuracy of flat secure sum protocols when progressive release is unused and offers an efficient accuracy–payload trade-off compared to distributed discrete-Gaussian and Skellam mechanisms.

This paper addresses a limitation of secure aggregation in federated learning: conventional protocols determine the precision of the aggregate at the time clients upload their updates. The authors introduce a progressive-resolution secure-aggregation functionality, in which clients upload once, but the server can later recover the aggregate at successively finer numerical resolutions. Rather than committing to a single precision level up front, the protocol supports a coarse-to-fine view of the same client-uploaded data, allowing the server to refine the aggregate after the initial communication round.

The key contribution is the formalization and construction of this new secure-aggregation primitive for federated learning. The central insight is that aggregate precision can be treated as an adaptive, post-upload parameter rather than a fixed property of the uploaded payload. Under the paper’s security assumptions, the server should be able to learn additional resolution of the aggregate without learning individual client updates, even though clients do not need to re-upload. This decouples communication from the final precision requirement and enables more flexible aggregation policies, such as stopping early when the current resolution is sufficient or refining the aggregate if higher accuracy is needed for training.

The work matters because communication efficiency and privacy are both first-order constraints in federated learning. Fixed-precision secure aggregation forces a tradeoff: clients may upload more precision than necessary, wasting bandwidth, or the server may need additional rounds to obtain a more accurate aggregate. Progressive-resolution secure aggregation offers a more practical middle ground, potentially reducing communication costs, supporting adaptive training procedures, and making secure federated learning easier to deploy in bandwidth-constrained or privacy-sensitive settings.