The challenges of scalability, robustness, and resilience to slow devices have posed significant obstacles to the effective and efficient implementation of Federated Learning (FL), a crucial technology for the emerging Artificial Intelligence of Things (AIoT). This paper proposes a solution to these challenges with the introduction of a Decentralized Hierarchical Asynchronous Federated Learning Scheme (DHA-FL). This scheme utilizes a hierarchical edge computing architecture, enabling a two-stage model aggregation paradigm that significantly enhances system scalability. To further enhance system robustness, decentralized asynchronous model aggregation is adopted among edge servers to prevent single node failures while mitigating the impact of slow devices or stragglers. Our experiments, conducted on a live wireless multi-hop IoT testbed, demonstrate that DHA-FL can achieve convergence in approximately half the time compared to the centralized hierarchical approach. Moreover, it enables an even more significant convergence speed-up (up to 8x) over the classic FedAvg baseline when dealing with stragglers.

Split-computing has recently emerged as a paradigm for offloading computation of visual analytics models from low-powered mobile devices to edge or cloud servers, by which the mobiles execute part of the model and compress and send the intermediate features, and the servers complete the remaining model computation. Prior feature compression approaches train different compression models and possibly visual analytics models to reach different target bit rates. We propose a scalable compression model that compresses the intermediate features of the YOLO object detection model into a layered bitstream, which can be easily adapted to meet the rate constraint of a dynamic network. Our approach achieves comparable rate-accuracy performance compared to prior non-scalable compression approaches over a large bitrate range.

How can we perform inference on data using cloud servers without leaking any information to them? The answer lies in Trained-MPC, an innovative approach to inference privacy that can be applied to deep learning models. It relies on a cluster of servers, each running a learning model, which are fed with the client data added with strong noise. The noise is independent of user data, but dependent across the servers. The variance of the noise is set to be large enough to make the information leakage to the servers negligible. The dependency among the noise of the queries allows the parameters of the models running on different servers to be trained such that the client can mitigate the contribution of the noises by combining the outputs of the servers, and recover the final result with high accuracy and with a minor computational effort. In other words, in the proposed method, we develop a multiparty computation (MPC) by training for a specific inference task while avoiding the extensive communication overhead that MPC entails. Simulation results demonstrate Trained-MPC resolves the tension between privacy and accuracy while avoiding the computational and communication load needed in cryptography schemes.

Asynchronous learning protocols have regained attention lately, especially in the Federated Learning (FL) setup, where slower clients can severely impede the learning process. Herein, we propose \texttt{AsyncDrop}, a novel asynchronous FL framework that utilizes dropout regularization to handle device heterogeneity in distributed settings. Overall, \texttt{AsyncDrop} achieves better performance compared to state of the art asynchronous methodologies, while resulting in less communication and training time overheads. We implement our approach and compare it against other asynchronous baselines, both by design and by adapting existing synchronous FL algorithms to asynchronous scenarios. Empirically, \texttt{AsyncDrop} reduces the communication cost and training time, while matching or improving the final test accuracy in diverse non-i.i.d. FL scenarios.

In recent years, federated learning has been widely studied to speed up various \textit{supervised} learning tasks at the wireless network edge under communication constraints. However, there is a lack of theoretical understanding as to whether similar speedups in sample complexity can be achieved for cooperative reinforcement learning (RL) problems subject to realistic communication models. To that end, we study a federated policy evaluation problem over wireless fading channels where, to update model parameters, a central server aggregates local temporal difference (TD) update directions from $N$ agents via analog over-the-air computation (OAC). We refer to this scheme as \texttt{OAC-FedTD} and provide a rigorous finite-time convergence analysis of its performance that accounts for linear function approximation, Markovian sampling, and channel-induced distortions and noise. Our analysis reveals the impact of the noisy fading channels on the convergence rate and establishes a linear convergence speedup w.r.t. the number of agents. As far as we are aware, this is the first non-asymptotic analysis of a cooperative RL setting under channel effects. Moreover, our proof leads to tighter bounds on the mixing time relative to existing work in federated RL (without channel effects); as such, it can be of independent interest.