A survey on federated learning

Knowledge-Based Systems, Год журнала: 2021, Номер 216, С. 106775 - 106775

Опубликована: Янв. 21, 2021

Язык: Английский

---

Energy Efficient Federated Learning Over Wireless Communication Networks

IEEE Transactions on Wireless Communications, Год журнала: 2020, Номер 20(3), С. 1935 - 1949

Опубликована: Ноя. 20, 2020

In this paper, the problem of energy efficient transmission and computation resource allocation for federated learning (FL) over wireless communication networks is investigated. considered model, each user exploits limited local computational resources to train a FL model with its collected data and, then, sends trained base station (BS) which aggregates broadcasts it back all users. Since involves an exchange between users BS, both latencies are determined by accuracy level. Meanwhile, due budget users, must be during process. This joint formulated as optimization whose goal minimize total consumption system under latency constraint. To solve problem, iterative algorithm proposed where, at every step, closed-form solutions time allocation, bandwidth power control, frequency, derived. requires initial feasible solution, we construct completion minimization bisection-based obtain optimal solution original problem. Numerical results show that algorithms can reduce up 59.5% compared conventional method.

Язык: Английский

---

Broadband Analog Aggregation for Low-Latency Federated Edge Learning

IEEE Transactions on Wireless Communications, Год журнала: 2019, Номер 19(1), С. 491 - 506

Опубликована: Окт. 15, 2019

To leverage rich data distributed at the network edge, a new machine-learning paradigm, called edge learning, has emerged where learning algorithms are deployed for providing intelligent services to mobile users. While computing speeds advancing rapidly, communication latency is becoming bottleneck of fast learning. address this issue, work focused on designing low-latency multi-access scheme end, we consider popular privacy-preserving framework, federated (FEEL), global AI-model an edge-server updated by aggregating (averaging) local models trained devices. It proposed that updates simultaneously transmitted devices over broadband channels should be analog aggregated “over-the-air” exploiting waveform-superposition property channel. Such aggregation (BAA) results in dramatical communication-latency reduction compared with conventional orthogonal access (i.e., OFDMA). In work, effects BAA performance quantified targeting single-cell random network. First, derive two tradeoffs between communication-and-learning metrics, which useful planning and optimization. The power control (“truncated channel inversion”) required tradeoff update-reliability [as measured receive signal-to-noise ratio (SNR)] expected update-truncation ratio. Consider scheduling cell-interior constrain path loss. This gives rise other SNR fraction exploited Next, latency-reduction respect traditional OFDMA proved scale almost linearly device population. Experiments based neural real dataset conducted corroborating theoretical results.

Язык: Английский

---

Machine Learning at the Wireless Edge: Distributed Stochastic Gradient Descent Over-the-Air

IEEE Transactions on Signal Processing, Год журнала: 2020, Номер 68, С. 2155 - 2169

Опубликована: Янв. 1, 2020

We study federated machine learning (ML) at the wireless edge, where power- and bandwidth-limited devices with local datasets carry out distributed stochastic gradient descent (DSGD) help of a parameter server (PS). Standard approaches assume separate computation communication, estimates are compressed transmitted to PS over orthogonal links. Following this digital approach, we introduce D-DSGD, in which employ quantization error accumulation, transmit their multiple access channel (MAC). then novel analog scheme, called A-DSGD, exploits additive nature MAC for over-the-air computation, provide convergence analysis approach. In first sparsify estimates, project them lower dimensional space imposed by available bandwidth. These projections sent directly without employing any code. Numerical results show that A-DSGD converges faster than D-DSGD thanks its more efficient use limited bandwidth natural alignment channel. The improvement is particularly compelling low power regimes. also illustrate classification problem that, robust bias data distribution across devices, while significantly outperforms other schemes literature. observe both perform better number showing ability harnessing edge devices.

Язык: Английский

---

A Survey on Federated Learning for Resource-Constrained IoT Devices

IEEE Internet of Things Journal, Год журнала: 2021, Номер 9(1), С. 1 - 24

Опубликована: Июль 6, 2021

Federated learning (FL) is a distributed machine strategy that generates global model by from multiple decentralized edge clients. FL enables on-device training, keeping the client's local data private, and further, updating based on updates. While methods offer several advantages, including scalability privacy, they assume there are available computational resources at each edge-device/client. However, Internet-of-Things (IoT)-enabled devices, e.g., robots, drone swarms, low-cost computing devices (e.g., Raspberry Pi), may have limited processing ability, low bandwidth power, or storage capacity. In this survey article, we propose to answer question: how train models for resource-constrained IoT devices? To end, first explore existing studies FL, relative assumptions implementation using their drawbacks. We then discuss challenges issues when applying an environment. highlight overview of provide comprehensive problem statements emerging challenges, particularly during within heterogeneous environments. Finally, point out future research directions scientists researchers who interested in working intersection

Язык: Английский

---

Latency Minimization for Intelligent Reflecting Surface Aided Mobile Edge Computing

IEEE Journal on Selected Areas in Communications, Год журнала: 2020, Номер 38(11), С. 2666 - 2682

Опубликована: Июль 3, 2020

Computation off-loading in mobile edge computing (MEC) systems constitutes an efficient paradigm of supporting resource-intensive applications on devices. However, the benefit MEC cannot be fully exploited, when communications link used for computational tasks is hostile. Fortunately, propagation-induced impairments may mitigated by intelligent reflecting surfaces (IRS), which are capable enhancing both spectral- and energy-efficiency. Specifically, IRS comprises controller a large number passive elements, each impose phase shift incident signal, thus collaboratively improving propagation environment. In this paper, beneficial role IRSs investigated systems, where single-antenna devices opt fraction their to node via multi-antenna access point with aid IRS. Pertinent latency-minimization problems formulated single-device multi-device scenarios, subject practical constraints imposed capability design. To solve problem, block coordinate descent (BCD) technique invoked decouple original problem into two subproblems, then settings alternatively optimized using low-complexity iterative algorithms. It demonstrated that our IRS-aided system significantly outperforming conventional operating without IRSs. Quantitatively, about 20 % latency reduction achieved over single cell 300 m radius 5 active devices, relying 5-antenna point.

Язык: Английский

---

Edge Artificial Intelligence for 6G: Vision, Enabling Technologies, and Applications

IEEE Journal on Selected Areas in Communications, Год журнала: 2021, Номер 40(1), С. 5 - 36

Опубликована: Ноя. 8, 2021

The thriving of artificial intelligence (AI) applications is driving the further evolution wireless networks. It has been envisioned that 6G will be transformative and revolutionize from "connected things" to intelligence". However, state-of-the-art deep learning big data analytics based AI systems require tremendous computation communication resources, causing significant latency, energy consumption, network congestion, privacy leakage in both training inference processes. By embedding model capabilities into edge, edge stands out as a disruptive technology for seamlessly integrate sensing, communication, computation, intelligence, thereby improving efficiency, effectiveness, privacy, security In this paper, we shall provide our vision scalable trustworthy with integrated design strategies decentralized machine models. New principles networks, service-driven resource allocation optimization methods, well holistic end-to-end system architecture support described. Standardization, software hardware platforms, application scenarios are also discussed facilitate industrialization commercialization systems.

Язык: Английский

---

Federated Machine Learning: Survey, Multi-Level Classification, Desirable Criteria and Future Directions in Communication and Networking Systems

IEEE Communications Surveys & Tutorials, Год журнала: 2021, Номер 23(2), С. 1342 - 1397

Опубликована: Янв. 1, 2021

The communication and networking field is hungry for machine learning decision-making solutions to replace the traditional model-driven approaches that proved be not rich enough seizing ever-growing complexity heterogeneity of modern systems in field. Traditional assume existence (cloud-based) central entities are charge processing data. Nonetheless, difficulty accessing private data, together with high cost transmitting raw data entity gave rise a decentralized approach called Federated Learning. main idea federated perform an on-device collaborative training single model without having share any third-party entity. Although few survey articles on already exist literature, motivation this stems from three essential observations. first one lack fine-grained multi-level classification where existing surveys base their only criterion or aspect. second observation focus some common challenges, but disregard other aspects such as reliable client selection, resource management service pricing. third explicit straightforward directives researchers help them design future overcome state-of-the-art research gaps. To address these points, we provide comprehensive tutorial its associated concepts, technologies approaches. We then highlight applications directions domain networking. Thereafter, three-level scheme categorizes literature based high-level challenge they tackle. Then, classify each into set specific low-level challenges foster better understanding topic. Finally, provide, within challenge, technique used particular challenge. For category desirable criteria aimed community innovative efficient solutions. best our knowledge, most terms techniques it covers presents.

Язык: Английский

---

Blockchain-Federated-Learning and Deep Learning Models for COVID-19 Detection Using CT Imaging

IEEE Sensors Journal, Год журнала: 2021, Номер 21(14), С. 16301 - 16314

Опубликована: Апрель 30, 2021

With the increase of COVID-19 cases worldwide, an effective way is required to diagnose patients. The primary problem in diagnosing patients shortage and reliability testing kits, due quick spread virus, medical practitioners are facing difficulty identifying positive cases. second real-world share data among hospitals globally while keeping view privacy concerns organizations. Building a collaborative model preserving major for training global deep learning model. This paper proposes framework that collects small amount from different sources (various hospitals) trains using blockchain based federated learning. Blockchain technology authenticates organization. First, we propose normalization technique deals with heterogeneity as gathered having kinds CT scanners. Secondly, use Capsule Network-based segmentation classification detect Thirdly, design method can collaboratively train privacy. Additionally, collected real-life data, which is, open research community. proposed utilize up-to-date improves recognition computed tomography (CT) images. Finally, our results demonstrate better performance

Язык: Английский

---

Distributed Learning in Wireless Networks: Recent Progress and Future Challenges

IEEE Journal on Selected Areas in Communications, Год журнала: 2021, Номер 39(12), С. 3579 - 3605

Опубликована: Окт. 6, 2021

The next-generation of wireless networks will enable many machine learning (ML) tools and applications to efficiently analyze various types data collected by edge devices for inference, autonomy, decision making purposes. However, due resource constraints, delay limitations, privacy challenges, cannot offload their entire datasets a cloud server centrally training ML models or inference To overcome these distributed techniques have been proposed as means collaboratively train without raw exchanges, thus reducing the communication overhead latency well improving privacy. deploying over faces several challenges including uncertain environment (e.g., dynamic channel interference), limited resources transmit power radio spectrum), hardware computational power). This paper provides comprehensive study how can be effectively deployed networks. We present detailed overview emerging paradigms, federated learning, distillation, multi-agent reinforcement learning. For each framework, we first introduce motivation it Then, literature review on use its efficient deployment. then an illustrative example show optimize improve performance. Finally, future research opportunities. In nutshell, this holistic set guidelines deploy broad range frameworks real-world

Язык: Английский

---