See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/344410435

Communicate to Learn at the Edge

Preprint · September 2020

DOI: 10.48550/arXiv.2009.13269

CITATIONS READS
0 174

6 authors, including:

Deniz Gündüz David Burth Kurka

Imperial College London Imperial College London
599 PUBLICATIONS 15,202 CITATIONS 33 PUBLICATIONS 1,778 CITATIONS

SEE PROFILE SEE PROFILE

Mikolaj Jankowski Mohammad Mohammadi Amiri

Imperial College London Rensselaer Polytechnic Institute
21 PUBLICATIONS 649 CITATIONS 60 PUBLICATIONS 2,593 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Mikolaj Jankowski on 29 March 2021.

The user has requested enhancement of the downloaded file.

 1

Communicate to Learn at the Edge

D. Gündüz1 , D. Burth Kurka1 , M. Jankowski1 ,
M. Mohammadi Amiri2 , E. Ozfatura1 , and S. Sreekumar3
1
Department of Electrical and Electronic Engineering, Imperial College London
2
Department of Electrical Engineering, Princeton University
3
Department of Electrical and Computer Engineering, Cornell University
arXiv:2009.13269v1 [eess.SP] 28 Sep 2020

Abstract

Bringing the success of modern machine learning (ML) techniques to mobile devices can enable
many new services and businesses, but also poses significant technical and research challenges. Two
factors that are critical for the success of ML algorithms are massive amounts of data and process-
ing power, both of which are plentiful, yet highly distributed at the network edge. Moreover, edge
devices are connected through bandwidth- and power-limited wireless links that suffer from noise,
time-variations, and interference. Information and coding theory have laid the foundations of reliable
and efficient communications in the presence of channel imperfections, whose application in modern
wireless networks have been a tremendous success. However, there is a clear disconnect between the
current coding and communication schemes, and the ML algorithms deployed at the network edge. In
this paper, we challenge the current approach that treats these problems separately, and argue for a joint
communication and learning paradigm for both the training and inference stages of edge learning.

I. M OTIVATION

Modern machine learning (ML) techniques have made tremendous advances in areas such
as machine vision, robotics, and natural language processing. Novel ML applications emerge
every day, ranging from autonomous driving and finance to marketing and healthcare – potential
applications are limitless. In parallel, the fifth generation (5G) of mobile technology promises
to connect billions of heterogeneous devices to the network edge, supporting new applications
and verticals under the banner of Internet of things (IoT). Edge devices will collect massive
amounts of data, opening up new avenues for ML applications. The prevalent approach for
the implementation of ML solutions on edge devices is to amass all the relevant data at a

This work was supported by the European Research Council (ERC) Starting Grant BEACON (grant agreement no. 677854).
 2

Figure 1: Distributed learning and inference at the wireless network edge.

cloud server, and train a powerful ML model using all the available data and processing power.
However, such a ‘centralized’ solution is not applicable in many cases. This might violate the
latency requirements of the underlying application, particularly in the inference stage; or, result
in the infringement of user privacy.Moreover, as the data volumes increase, limited bandwidth
and energy resources of IoT devices will become a bottleneck. For example, an autonomous car
generates 5 to 20 terabytes of data per day. This is a particular challenge when the ‘information
density’ of the collected data is low, i.e., large volumes of data with only limited relevant
information for the underlying learning task.
To meet the requirements of most IoT applications, the ‘intelligence’ should move from the
centralized cloud to the network edge. However, both data and processing power, the essential
constituents of machine intelligence, are highly distributed at the edge. As a result, commu-
nication becomes key to an intelligent network edge, and potential solutions must allow
edge devices not only to share their data but also computational resources in a seamless and
efficient manner. We can argue that the current success of ML, thanks to the tremendous
increase in computational power, is similar to the ‘great leap forward’ in human evolution,
which led to the development of human brain thanks to a favorable mutation. Continuing with
this analogy, next big revolution in ML is likely to arrive thanks to the efficient orchestration and
collaboration among intelligent devices, similarly to the impact of language in human history,
which tremendously accelerated the advancement of our civilization by allowing humans to share
information, experience, and intelligence.
 3

A. The Communication Challenge

Communication bottleneck in ML has been acknowledged in the literature; yet, most current
approaches treat communication links as rate-limited ideal bit pipes. However, wireless links
introduce errors due to noise and channel fading, and error-free operation is either impossible,
or would result in significant delays. This is particularly prominent at the network edge, where
bandwidth- and power-limited IoT devices share the same wireless medium, also creating inter-
ference to each other. Moreover, when information moves across a network, privacy and security
concerns arise, exacerbated at the edge due to the vulnerability of individual devices and the
broadcast nature of wireless transmissions.
After decades of research, communication engineers have designed highly advanced coding
and communication techniques that can mitigate channel imperfections and create reliable links
among wireless devices; however, reducing the communication among edge devices to a network
of ideal bit pipes has the following limitations: 1) communication protocols that enable such
reliable links introduce significant overheads and delays, which are not acceptable for many ML
applications; 2) such levels of reliability at the link level may not be required for some ML
applications, resulting in inefficient resource management; 3) most communication protocols
are designed to reduce or remove interference, which may not be desired in some distributed
ML applications. To overcome these limitations, we need to reconsider physical layer and
networking solutions taking into account the limitations and requirements of the underlying
ML applications.
Information and coding theory have laid the foundations of reliable, efficient and secure
communication in the presence of channel imperfections and interference, whose application in
modern wireless networks have been a tremendous success. While the fundamental information
theoretic ideas and coding theoretic tools can play an important role in enabling fully distributed
learning across distributed heterogeneous edge devices, many of the existing concepts and tech-
niques are not relevant for ML applications, whose communication requirements and constraints
(latency, reliability, security, privacy, etc.) are fundamentally different from the type of traffic
current networks are designed for. Moreover, as we will try to show in this paper, we cannot
overcome these limitations by a simple ‘cross-layer’ approach, i.e., by tuning the parameters of
existing communication protocols. There is a clear disconnect between the current coding and
communication techniques, and the ML algorithms and architectures that must be deployed at
 4

Figure 2: DNNs at the network edge.

the network edge, and we need a fundamentally new paradigm of coding, communication
and networking with ML applications in mind.
Next, we present the challenges in achieving a fully distributed edge intelligence across het-
erogeneous agents communicating over imperfect wireless channels. We will treat the inference
and training phases of ML algorithms separately as they have distinct reliability and latency
requirements.

II. D ISTRIBUTED I NFERENCE

Inference refers to applying a trained model on a new data sample to make a prediction.
Although inference tasks require much less computational resources compared to training, they
typically impose more strict latency constraints. For example, in self-driving cars (see Fig. 1),
immediate detection of obstacles is critical to avoid accidents. A powerful deep neural network
(DNN) model can be pre-trained and deployed for this task. However, it is often not possible to
carry out inference locally at a single device, as decisions may rely on data (e.g., background
and terrain information) available at an edge server, or on signals from other cars; or the device
gathering the data (e.g., a bike) may not have the necessary processing capability. Communication
becomes indispensable in such scenarios, and we need to guarantee that inference can still be
accomplished within the accuracy and latency constraints of the underlying application.
Fundamental limits. As a first step towards understanding the fundamental limits of statistical
inference over noisy channels, a distributed binary hypothesis testing (HT) problem is studied
in [1]. Consider two devices with their local observations. One of the devices (e.g., the car
in Fig. 1), called the observer, conveys some information about its observations to the other
one, called the decision maker (e.g., the edge server in Fig. 1), over a noisy channel. The
 5

decision maker has to make a decision on the joint distribution of the observations of the two
devices. Since the observer has access only to its own observations, it cannot make a local
decision no matter how much processing power it has; instead it must convey some features of
its observations to help the decision maker to make the correct decision. The question here is
whether the features and the channel code to transmit them can be designed separately. If the
goal were to transmit the samples at the observer with the minimal average distortion (under any
additive finite distortion measure), according to Shannon’s separation theorem the compression
and channel coding tasks can be carried out separately and without loss of optimality, in the
limit of infinite blocklength. However, it is shown in [1] that the optimality of separation breaks
down in the remote HT problem, as the goal here is to decide on the joint distribution with
minimal error probability. While this result shows that communication and inference cannot be
separated even in the asymptotic limit (without loss of optimality), how a joint scheme should be
designed in practice is a vastly unexplored research direction with great potential in future edge
inference applications. Next, we provide several practical examples of edge inference problems,
and illustrate how jointly treating communication and inference can help improve both the speed
and the accuracy of the inference task.
Edge Inference with DNNs. DNNs achieve the state-of-the-art performance in most ML
tasks. In distributed inference across mobile devices and edge servers, a common approach is
to partition a pre-trained DNN baseline between the devices and the edge server depending on
the former’s computational capabilities (see Fig. 2) [2]. Conventional approaches abstract out
the wireless channel as an error-free ideal bit-pipe, and focus only on the feature compression
problem, ignoring the potential impacts of communication in terms of delay, complexity, and
reliability. However, lossy transmission of feature vectors over a wireless channel is a joint
source-channel coding (JSCC) problem, and separation is known to be suboptimal under strict
latency constraints imposed by inference problems.
While JSCC has long been studied, mainly for image and video transmission, these works
mostly took a model-driven approach exploiting particular properties of the underlying source
and channel statistics. Recently, an alternative fully data-driven DNN-based scheme, called
DeepJSCC, has been introduced in [3]. DeepJSCC not only beats digital alternatives for image
transmission (e.g., BPG image compression + LDPC channel coding), but also provides ‘graceful
degradation’ with channel quality, making it ideal for IoT applications, where accurate channel
estimation is often not possible. DeepJSCC also reduces the coding/decoding delay compared to
 6

0.50

0.45

0.40

Top-1 accuracy
0.35

0.30

0.25 Re-ID baseline

(w/o channel)
0.20 End-to-end
analog approach
0.15 Digital approach
(capacity-achieving)
0.10
10 5 0 5 10 15
Channel SNR [dB]

Figure 3: Accuracy vs. channel SNR for remote person re-ID over an AWGN channel.

conventional digital schemes more than 5 times on a CPU, and more than 10 times on a GPU.
As opposed to conventional digital schemes, DeepJSCC can easily adapt to specific information
source or channel statistics through training, e.g., landscape images transmitted from a drone
or a satellite. This makes DeepJSCC especially attractive for edge inference as we do not have
compression codes designed for feature vectors, whose statistics would change from application
to application.
A practical edge inference problem is studied in [4], where the image of a person captured by
a remote camera is to be identified within a database available at an edge server, called the re-
identification (re-ID) problem. Here, the camera cannot make a local decision as it does not have
access to the database. In [4], two approaches are proposed, both employing DNNs for remote
inference: a task-oriented DNN-based compression scheme for digital transmission and a DNN-
based analog JSCC approach, à la DeepJSCC. These schemes are compared in Fig. 3 in terms
of top-1 identification accuracy when only 128 real symbols are transmitted over an additive
white Gaussian noise (AWGN) channel. We observe that the analog approach, which maps the
feature vectors directly to channel inputs (no explicit compression or channel coding), performs
significantly better, achieving the baseline performance around a channel signal-to-noise-ratio
(SNR) of approximately 8 dB. We highlight that the conventional scheme of transmitting the
query images with the best possible quality (ignoring the learning task), and then applying the
re-ID baseline on the reconstructed image is not included as it would require much higher SNR
 7

values to achieve a comparable performance. This result shows that separating communication
from inference at the edge can be highly suboptimal. While joint design can offer significant
performance gains, it brings about new challenges and requires novel coding and communication
paradigms, including the extension of the proposed edge inference approach to time-varying
and/or non-Gaussian channels, and to multi-antenna and multi-user networks.
In the inference stage, the challenge is to convey the most relevant information about the data
samples to the decision maker to achieve the desired level of accuracy within the constraints of
the edge network. The results above show that the channel characteristics must be taken into
account during the training stage, rather than being abstracted out, and effectively, we learn how
to communicate and infer jointly. In this section, we have assumed that the DNNs are trained
centrally, and then deployed at the edge devices, assuming the availability of sufficient training
data and an accurate model of the wireless communication channels. We focus on the training
stage in the next section.

III. D ISTRIBUTED T RAINING

Training is particularly challenging at the network edge due to the distributed nature of both
the data and the processing power. Below, we will first address the scenario in which an edge
device with its own dataset employs the computational resources of multiple edge servers to
speed up training (see Fig. 4a). Later, we will consider the scenario when data is also distributed
(see Fig. 4b).
In the training stage of a standard ML problem, the goal is to optimize the model param-
eters over a training dataset with respect to an application specific empirical loss function.
This optimization problem is typically solved by stochastic gradient descent (SGD), iteratively
updating the parameter vector along the estimated gradient descent direction. This algorithm is
highly parallelizable, allowing distributed and parallel implementation. When the dataset is large,
distributed SGD across multiple edge servers can be utilized to reduce the training time. The
dataset can be divided into non-overlapping subsets, each given to a different server. At each
iteration of the gradient descent algorithm, the user broadcasts the current model parameters
to all the servers. Each server computes a partial gradient based only on its local dataset, and
returns the result to the master. The master waits to receive partial gradients from all the servers
in order to aggregate them and obtain the full gradient. In this implementation, however, due
to synchronised updates the completion time of each iteration is constrained by the straggling
 8

(a) Distributed training with centralized data.

(b) FEEL with distributed data.

Figure 4: Distributed training at the edge.

server(s), where the straggling may be due to failing hardware, contention in the network, or
even channel outages if the training is carried out at the wireless edge.
Straggling servers can be treated as ‘erasures’, and using ideas from coding theory, redundant
computations can be introduced to efficiently compensate for erasures [5], [6]. This can help
reduce the recovery threshold, the minimum number of responsive servers required to complete
the computation task, e.g., computing a sufficiently accurate gradient estimate. However, this may
require coding the data before offloading to the servers [5], or coding the results of computations
at each server [6], and eventually decoding these responses by the user, which introduce additional
 9

complexity and delays. Despite the significant research efforts in recent years, optimal coding
schemes remain elusive, and there is no comprehensive analysis of end-to-end latency that take
into account the communication, coding, and computing delays.
Moreover, most of the existing techniques suffer from two main drawbacks: the recovery
threshold can be reduced by increasing the redundancy; yet, the servers may end up executing
more computations than required due to an inaccurate prediction of the straggling behaviour,
resulting in over-computation. Also, most of the existing solutions are designed for persistent
stragglers, and partial computations carried out by stragglers are discarded, resulting in under-
utilization of the computational resources. To overcome these limitations, each server can be
allowed to send multiple messages during each training iteration [7], each corresponding to
partial computations. This approach will provide additional flexibility for straggler mitigation,
resulting in a trade-off between the amount of communication and computation. We highlight that
the real performance indicator for these schemes is the average completion time of training, which
requires the joint design of the underlying communication protocol and the coded computing
scheme employed.
Private and secure distributed computation. Distributed training also introduces privacy and
security challenges. Malicious servers can inject false data, while honest but curious servers can
exploit user data for purposes beyond computation. Coded computing, in particular polynomial
codes, can provide security and privacy guarantees in addition to straggler mitigation by deliv-
ering coded data samples to the computing servers [8], but the optimal trade-off between the
required communication bandwidth between the user and the servers, and the privacy/ security
guarantees (in terms of the number of colliding servers) remains an open challenge.

IV. F EDERATED EDGE LEARNING (FEEL)

When multiple edge devices with their own local datasets collaborate to train a joint model,
devices may not want to offload their data due to privacy concerns. Yet, unlike in distributed
training, data samples at different devices cannot be coded to provide privacy. Federated learning
(FL) has been introduced by Google to enable collaborative training without sharing local datasets
[9], typically orchestrated by a parameter server (PS) (see Fig. 4b). In FL, the PS broadcasts
a global model to the devices. Each device runs SGD locally using the current global model.
Device updates are aggregated at the PS, and used to update the global model. Communication,
again, is a major challenge due to the bandwidth and power limitations of devices. To reduce
 10

the communication load, random subsets of devices are selected at each round, and local models
are communicated after several local SGD updates. Another approach is to reduce the size of
the messages communicated between the devices and the PS through compression. This is yet
another research challenge where the extensive knowledge in information and coding theory for
data compression can make an impact. While initial works have focused on rather simple scalar
quantization and sparsification techniques [10], more advanced vector quantization and temporal
coding tools exploiting correlations across gradient dimensions or multiple iterations can further
reduce the communication load. But, the complexity of such tools must be carefully balanced
with the potential gains.
In FEEL, we assume that the training takes place at the network edge across wireless devices
within physical proximity; therefore, communication from edge devices to the PS will be limited
by the power and bandwidth constraints, interference among devices, and time-varying channel
fading. When the model size is relatively small compared to the size of the dataset, exchanging
model parameters rather than data provides another advantage of FEEL. Still, allocation and
optimization of channel resources among devices will be essential to improve the learning per-
formance. On the other hand, conventional solutions that maximize throughput do not necessarily
translate into better accuracy or faster convergence in FEEL [11], [12]. Moreover, conventional
measures based on number of iterations may not be relevant in FEEL, as the wall clock time
depend hugely on the communication protocol [12]. Optimizing the communication protocols for
FEEL poses many interesting research challenges; however, most current approaches, motivated
by conventional communication systems, consider orthogonal resource allocation with the aim
of minimizing interference.
Interference can be a bliss. In the uplink transmission from the devices, the PS is interested
only in the average of the local models. Hence, rather than transmitting individual updates in
an orthogonal fashion, signal superposition property of the wireless medium can be exploited
to directly convey the sum of the local parameters through over-the-air computation [13], [14].
This is achieved by all the devices synchronously transmitting their model updates in an uncoded
‘analog’ fashion, which are superposed by the channel.
Uplink transmission of local model updates in FEEL is a distributed computation problem, for
which there is no separation theorem even when the sources are independent. Model updates at
different devices are highly compressible, and are often correlated. Hence, when model updates
are conveyed through digital communication, model compression can be used to adapt to the
 11

0.9

0.8

0.7
Test accuracy 0.6

0.5
Federated averaging (IID)
0.4 Analog over-the-air averaging (IID)
Digital transmission with quantization (IID)
Federated averaging (non-IID)
0.3 Analog over-the-air averaging (non-IID)
Digital transmission with quantization (non-IID)
0.2
0 50 100 150 200 250 300 350 400
Iteration count, t

Figure 5: Test accuracy of FEEL for MNIST classification with IID and non-IID data distribu-
tions.

limited channel resources available to each device [10]. In analog transmission, however, even
though all the devices transmit over the same channel resources, the required bandwidth can be
fairly large. Some state-of-the-art models include tens of millions of parameters, whereas 1 LTE
frame of 5MHz bandwidth and 10ms duration can carry only 6K complex symbols. In [13],
sparsification of model updates is proposed followed by linear projection with a pseudo-random
Gaussian matrix. This novel approach serves as an analog compression technique, and reliable
reconstruction can be achieved by approximate message passing at the PS. In Fig. 5, we compare
digital and analog schemes for the MNIST classification task over a Gaussian multiple access
channel. In the IID case, local datasets are chosen randomly from the whole training dataset;
whereas in the non-IID case each device has samples from only two classes. We see that over-
the-air computation provides significant gains in both the final accuracy and the convergence
speed. Over-the-air computation allows scheduling more devices within the same time constraint,
which provides variance reduction in updates, and better robustness against the channel noise
[13]. This is yet another example, where a joint design of the communication and learning
algorithms is essential.
We remark that over-the-air computation assumes symbol-level synchronization among the
 12

participating devices. In practice, this can be achieved through a synchronization channel, e.g.,
timing advance in LTE systems, resulting in a trade-off between the overall performance and
the resources dedicated to synchronization, which is an interesting research direction to fully
evaluate the potential benefits of over-the-air computation for FEEL.
Privacy in FEEL. Although FL has been introduced as a privacy-aware solution for collab-
orative learning, it is known to be vulnerable to membership as well as reconstruction attacks
solely using the gradient information [15]. Although differential privacy can be achieved by
introducing noise into the gradients transmitted by the devices, this typically requires adding
significant amount of noise, making the model hard to converge. On the other hand, in FEEL,
there is inherent noise and interference in the channel, which can be exploited to increase the
security and privacy of the system through purely physical layer techniques. This opens up a
new type of physical layer security/ privacy framework for FEEL applications.

V. D ISCUSSION AND C ONCLUSIONS

Communication will play an essential role in employing ML tools at the network edge. Current
approaches to communication-efficient distributed ML ignore the physical layer, and assume
error and delay-free ideal links. This approach presumes a communication protocol, designed
independently of the learning task, taking care of channel imperfections. In this paper, we have
argued through references to recent theoretical results and practical implementations that such
a separate architecture can be highly suboptimal, and a novel joint communication and learning
framework is essential in approaching the fundamental limits of distributed learning. This calls
for a new research paradigm integrating coding and communication theoretic ideas within the
design of ML algorithms at the network edge. We have shown that the benefits of such a
joint design paradigm can be significant for edge inference, both to boost the final performance
and to meet the stringent delay constraints. Training is more computation intensive compared
to inference; hence, computation and communication delays in training must be optimized
jointly. Furthermore, heterogeneity of edge servers may result in additional bottlenecks due to
stragglers. Coding can be used both to reduce the computation delays and to mitigate stragglers.
Moreover, each iteration of the training process can be considered as a distributed computation
problem, which renders throughput-maximizing conventional communication protocols obsolete,
and requires the design of novel communication protocols and coding schemes. Since training
is carried out in many (imperfect) iterations, we can relax some of the constraints of traditional
 13

coding and communication schemes (reliability, synchronization, power control, etc), resulting in
novel communication problems. Finally, taking into account the physical layer channel charac-
teristics can allow exploiting coding and communication theoretic tools to provide fundamental
information theoretic privacy and security guarantees for both inference and training at the edge.
Each of these perspectives and challenges open up new research problems in this exciting new
research area exploring the connections between communication and learning.

R EFERENCES

[1] S. Sreekumar and D. Gündüz, “Distributed hypothesis testing over discrete memoryless channels,” IEEE Transactions on
Information Theory, vol. 66, no. 4, pp. 2044–2066, 2020.
[2] A. Eshratifar, M. Abrishami, and M. Pedram, “JointDNN: An efficient training and inference engine for intelligent mobile
cloud computing services,” IEEE Trans. on Mobile Computing, 2019.
[3] E. Bourtsoulatze, D. Kurka, and D. Gündüz, “Deep joint source-channel coding for wireless image transmission,” IEEE
Trans. on Cogn. Comms. and Networking, vol. 5, no. 3, pp. 567–579, Sep. 2019.
[4] M. Jankowski, D. Gündüz, and K. Mikolajczyk, “Deep joint source-channel coding for wireless image retrieval,” in IEEE
ICASSP, 2020, pp. 5070–5074.
[5] K. Lee, M. Lam, R. Pedarsani, D. Papailiopoulos, and K. Ramchandran, “Speeding up distributed machine learning using
codes,” IEEE Trans. Inf. Theory, vol. 64, no. 3, pp. 1514–1529, Mar. 2018.
[6] R. Tandon, Q. Lei, A.G. Dimakis, and N. Karampatziakis, “Gradient coding: Avoiding stragglers in distributed learning,”
in Int’l Conf. on Machine Learning, Aug. 2017, pp. 3368–3376.
[7] E. Ozfatura, S. Ulukus, and D. Gunduz, “Straggler-aware distributed learning: Communication-computation latency trade-
off,” Entropy: Special Issue on The Interplay Between Storage, Computing, and Communications from An Information-
Theoretic Perspective, vol. 22, no. 5, May 2020.
[8] Q. Yu, S. Li, N. Raviv, S. Kalan, M. Soltanolkotabi, and S. Avestimehr, “Lagrange coded computing: Optimal design for
resiliency, security, and privacy,” in Proc. of Machine Learning Research, Apr. 2019.
[9] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B.A. y Arcas, “Communication-efficient learning of deep networks
from decentralized data,” in Proc. Int’l Conf. on Artificial Intelligence and Stat., Apr. 2017, pp. 1273–1282.
[10] D. Alistarh, D. Grubic, J. Li, R. Tomioka, and M. Vojnovic, “QSGD: Communication-efficient SGD via gradient
quantization and encoding,” in Advances in Neural Inform. Proc. Systems, 2017.
[11] H.H. Yang, Z. Liu, T. Quek, and V. Poor, “Scheduling policies for federated learning in wireless networks,” IEEE Trans.
on Comms., vol. 68, no. 1, pp. 317–333, Jan 2020.
[12] N.H. Tran, W. Bao, A. Zomaya, M.N.H. Nguyen, and C.S. Hong, “Federated learning over wireless networks: Optimization
model design and analysis,” in IEEE INFOCOM Conference on Computer Communications, 2019, pp. 1387–1395.
[13] M. Mohammadi Amiri and D. Gunduz, “Federated learning over wireless fading channels,” IEEE Transactions on Wireless
Communications, vol. 19, no. 5, pp. 3546–3557, 2020.
[14] ——, “Machine learning at the wireless edge: Distributed stochastic gradient descent over-the-air,” IEEE Transactions on
Signal Processing, vol. 68, pp. 2155–2169, 2020.
[15] L. Zhu, Z. Liu, and S. Han, “Deep leakage from gradients,” in Advances in Neural Information Processing Systems 32,
2019, pp. 14 774–14 784.

View publication stats