A DAG-based asynchronous federated learning framework with byzantine-resilient privacy preservation for edge IoT

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This paper addresses the trilemma of efficiency, Byzantine resilience, and differential privacy in asynchronous federated learning (FL) for edge IoT. The proposed DAG-AF 2 L framework is a fully decentralized architecture built on a directed acyclic graph (DAG) ledger that enables concurrent, non-blocking model updates without a central server. The framework jointly optimizes staleness-aware aggregation, multi-dimensional trust evaluation with graph-coupled reputation propagation, and trust-aware personalized differential privacy. A unified convergence analysis explicitly quantifies the effects of Byzantine attacks, DP noise, and asynchronous staleness, revealing a trinity tradeoff among robustness, privacy, and efficiency. Extensive experiments on three datasets under four attack types demonstrate that DAG-AF 2 L consistently outperforms state-of-the-art baselines in convergence, robustness, and privacy-utility tradeoff. The joint co-design eliminates destructive interference between DP noise and Byzantine filtering, validated through comprehensive detection metrics, and the empirical tradeoff surface corroborates the theoretical analysis. Deployment on an edge testbed confirms its practicality on resource-constrained devices. These results demonstrate that secure and private asynchronous FL can be realized through principled co-design on a DAG substrate.