With an increasing number of smart devices like internet of things devices deployed in the field, offloading training of neural networks (NNs) to a central server becomes more and more infeasible. Recent efforts to improve users’ privacy have led to on-device learning emerging as an alternative. However, a model trained only on a single device, using only local data, is unlikely to reach a high accuracy. Federated learning (FL) has been introduced as a solution, offering a privacy-preserving tradeoff between communication overhead and model accuracy by sharing knowledge between devices but disclosing the devices’ private data. The applicability and the benefit of applying baseline FL are, however, limited in many relevant use cases due to the heterogeneity present in such environments. In this survey, we outline the heterogeneity challenges FL has to overcome to be widely applicable in real-world applications. We especially focus on the aspect of computation heterogeneity among the participating devices and provide a comprehensive overview of recent works on heterogeneity-aware FL. We discuss two groups: works that adapt the NN architecture and works that approach heterogeneity on a system level, covering Federated Averaging, distillation, and split learning–based approaches, as well as synchronous and asynchronous aggregation schemes.

In Software-Defined Networking (SDN), the typical switch-controller cycle, from generating a network event notification at the controller until the flow rules are installed at the switches, is not an instantaneous activity. Our measurement results show that this has serious implications on the performance of flow setup procedure, specifically for larger networks: we observe that, even with software switches, the flow setup latency for networks of around 500 switches is in order of 50 milliseconds, with 99th percentile exhibiting 10 times higher latencies. To reduce both the latency and the variance of the flow setup, we propose path aggregation strategies, which turn the network into a set of pre-configured pipes that connect any pair of nodes. Our approach radically simplifies the flow setup procedure by minimizing the set of switches to be updated for new userinitiated flows to a constant number. We implement our solution in our testbed and study its performance through measurements. The results show that, in similar settings, it reduces the median and 99-percentile latencies to 5.9ms and 7ms, respectively, significantly improving the performance, especially in the tail.

The Intelligent Transportation System (ITS) environment is known to be dynamic and distributed, where participants (vehicle users, operators, etc.) have multiple, changing and possibly conflicting objectives. Although Reinforcement Learning (RL) algorithms are commonly applied to optimize ITS applications such as resource management and offloading, most RL algorithms focus on single objectives. In many situations, converting a multi-objective problem into a single-objective one is impossible, intractable or insufficient, making such RL algorithms inapplicable. We propose a multi-objective, multi-agent reinforcement learning (MARL) algorithm with high learning efficiency and low computational requirements, which automatically triggers adaptive few-shot learning in a dynamic, distributed and noisy environment with sparse and delayed reward. We test our algorithm in an ITS environment with edge cloud computing. Empirical results show that the algorithm is quick to adapt to new environments and performs better in all individual and system metrics compared to the state-of-the-art benchmark. Our algorithm also addresses various practical concerns with its modularized and asynchronous online training method. In addition to the cloud simulation, we test our algorithm on a single-board computer and show that it can make inference in 6 milliseconds.