Giskard: A Byzantine-Robust and Privacy-Preserving Aggregation Protocol for Large-Scale Decentralized Learning

The paper proposes Giskard, a protocol that simultaneously achieves privacy protection and Byzantine-robust aggregation in large-scale decentralized learning. In decentralized learning, clients keep data locally and share model parameters or gradients, making them vulnerable to privacy attacks and model poisoning attacks; client-side differential privacy amplifies the variance of honest updates and weakens Byzantine-robust aggregation, while previous schemes based on secure multi-party computation scale poorly due to all-to-all communication or reliance on a small number of computing parties. Giskard organizes n parties into a committee tree of size O(log n), and within each committee uses secure multi-party computation to perform a committee-adapted secure quantile search, obtaining a differentially private per-coordinate approximate median. The authors theoretically prove its security and privacy properties and validate it in experiments with up to one million parties, reducing per-party communication complexity by nearly three orders of magnitude compared with the closest competing scheme, while maintaining comparable robustness under up to n/4 Byzantine parties.