2024 International Conference on Advances in Computing, Communication,
Electrical, and Smart Systems (iCACCESS) 8-9 March, Dhaka, Bangladesh
CNN-Based End-to-End Deeper Autoencoders for
Physical Layer of Wireless Communication System
Jannatul Ferdous
Department of ETE
RUET
Rajshahi, Bangladesh
jannatulferdous1754@gmail.com
Md. Aslam Mollah
Department of ETE
RUET
Rajshahi, Bangladesh
ruet10aslam@gmail.com
Afiquer Rahman
Department of ETE
RUET
Rajshahi, Bangladesh
afiq.ruet18@gmail.com
Abstract—In this paper, a deeper autoencoder is proposed,
which is composed of systematic and strategic convolutional
neural network (CNN) layers. The proposed autoencoder can
function under various conditions of wireless communication sys-
tems, such as AWGN, Rayleigh, and other fading environments.
This model with perfect channel state information is compatible
with different channel use, data rates, throughput, and variations
in block length. The proposed system is an end-to-end task that
can optimize the transceiver’s components jointly and inherit
the sophisticated features of convolutional neural networks, such
as fast training convergence, generalization, classification, and
feature learning. In addition, simulation results illustrate that
the proposed model outperforms the existing models in terms of
block error rate. Another autoencoder is proposed in this paper,
where VGG-16 architecture is implemented before feeding to the
convolutional layer at the decoder side. A comparison of block
error rate performance shows that this scheme performs better
than the human expertise models.
Index Terms—Autoencoder, convolutional neural network,
wireless communication, channel state information, end-to-end
task, VGG-16.
I. I NTRODUCTION
In today’s digitized world, wireless networks and related
services have evolved into vital building blocks that have
revolutionized how we live, work, and interact. The growing
popularity of wireless technology makes wireless communi-
cation networks increasingly crowded, making efficient use
of wireless resources crucial. Therefore, the management and
allocation of resources, such as the transmission power, must
be done effectively to maximize the efficiency of wireless
communications. The communications field is particularly rich
in specialized knowledge derived from information theory,
statistics, and solid mathematical modeling proficient in chan-
nel modeling, optimal signaling, and detection algorithms for
consistent data transfer to compensate for numerous hardware
issues, etc., particularly for the physical layer. Ensuring opti-
mal communication between the transmitter and receiver end
despite the noise and distortions produced separately.
One of the latest innovations in the field of artificial intel-
ligence and machine learning is deep Learning (DL). Com-
puter vision, digital image processing, and natural language
processing have already experienced significant advancements
thanks to DL methodologies [1]. In recent years, DL has
been implemented in physical-layer communications systems
such as channel estimation [2] - [4], modulation classification
[5], and signal detection [6], [7]. It has also demonstrated
impressive outcomes, which are close to or even better than
the conventional systems.
At the physical layer, to decrease system complexity and
concentrate on the explicit function of each processing block,
the traditional communication system uses the divide-and-
conquer strategy relying on mathematical models. It divides
the transmitter and receiver into subtasks, such as modula-
tion, equalization, source coding, and channel coding. This
model, whose benefit is to enable component-by-component
optimization, has proven to be particularly effective in various
communications systems, including wireless communications.
On the other hand, complex systems, which experience unde-
sirable effects, are challenging to be modeled mathematically.
Therefore, a more adaptive and flexible paradigm may be
required to address these issues. This has led one creative
viewpoint to consider wireless communications as a self-
learned end-to-end autoencoder [8] that jointly optimizes
transmitter and receiver components so that learned weights
of neural networks facilitate encoding and decoding.
In terms of communications, the main objective of an autoen-
coder is to detect representations of the inputs (transmitted
signals) at some intermediate layer that are resilient concerning
the channel impairments mapping (e.g., noise, fading, and
distortion), allowing reconstruction at the output (received
signal) with a low probability of error. The self-learning
system can be considered weights learned from an optimized
neural network using end-to-end loss functions.
In the pioneering work of [9], a CNN-based end-to-end
autoencoder was proposed where the communication system
performed intelligently. However, the idea of using convo-
lutional layers as the fundamental blocks of autoencoder
(AE) for the communication system was introduced by the
979-8-3503-5028-9/24/$31.00 ©2024 IEEE
2024 International Conference on Advances in Computing, Communication, Electrical, and Smart Systems (iCACCESS) | 979-8-3503-5028-9/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICACCESS61735.2024.10499465
Authorized licensed use limited to: Rajshahi University Of Engineering and Technology. Downloaded on April 29,2024 at 18:35:48 UTC from IEEE Xplore. Restrictions apply.