Giskard : Byzantine Robust and Confidential Aggregation for Large-Scale Decentralized Learning

A new protocol named Giskard enables confidential and Byzantine-robust aggregation for decentralized machine learning, organizing participants into a tree of committees to reduce communication costs while tolerating adversarial behavior, according to research posted on arXiv [1]. The protocol, detailed in a paper submitted June 17, 2026, addresses a persistent tension in decentralized learning: the need to hide model updates for privacy while simultaneously inspecting them to detect malicious contributions [1]. Giskard organizes n parties into a tree of committees of size O(log n) and evaluates a coordinate-wise approximate median through a committee-adapted distributed binary search over the value domain, using BGW-style secure multi-party computation within each committee [1][2]. The authors argue that existing solutions scale poorly, either requiring all-to-all communication or concentrating computation on a small subset whose load grows with network size [1]. Giskard instead distributes the aggregation workload across multiple committees, so no party reveals its individual gradient vector in the clear [2][3]. The protocol is designed to be UC-secure and tolerates up to n/4 Byzantine nodes, matching the robustness threshold of coordinate-wise median [2][3]. Experiments involved up to one million participants, and the paper reports that Giskard reduces per-party communication complexity asymptotically while maintaining model utility comparable to its closest competitors under adversarial conditions [1][2]. A key technical insight is reformulating the coordinate-wise median as a binary-search problem, reducing robust aggregation to a sequence of secure counting operations rather than relying on generic sorting or comparison circuits [2][3]. Other recent work has explored similar goals. The SecureDL protocol, for instance, uses cosine similarity and normalization coupled with secret sharing to detect Byzantine updates while preserving privacy, and was evaluated on MNIST, Fashion-MNIST, SVHN, and CIFAR-10 datasets [4]. SecureDL demonstrated effectiveness even when 80 percent of clients were Byzantine, though its architecture differs from Giskard's committee-based approach [4]. The Giskard paper appears on arXiv, the open-access repository that hosts preprints across mathematics, physics, computer science, and related fields [8]. As of November 2024, arXiv was receiving roughly 24,000 submissions per month and had surpassed two million total articles by the end of 2021 [8].