International Journal of

INTELLIGENT SYSTEMS AND APPLICATIONS IN

ENGINEERING
ISSN:2147-67992147-6799 www.ijisae.org Original Research Paper

Deep Learning Algorithm Using Densenet to Enable Big Data Analytics

in Large WiFi Systems
1
Doradla Bharadwaja, 2R. Gayathri, 3D. Sugumar, 4D. R. Denslin Brabin, 5Faiz Akram, 6Mohd. Javeed
Mehdi
Submitted: 03/10/2023 Revised: 22/11/2023 Accepted: 03/12/2023
Abstract: The increasing mobile device and unceasing traffic demand enables the deployment of large-scale WiFi systems that offers
indoor coverage and high-speed connectivity. The large-scale deployment of WiFi system is an on-going research in wireless system due
to its challenging heterogeneous nature of access points. Such access points undergo rapid challenges due to traffic conditions and traffic
consumptions with rapidly increasing input data. On other hand, massive connection with heavy traffic laden from the WiFi devices
poses increased pressure on backhaul network and reduces the Quality of Service by the users. We have developed using DenseNets that
reduces the backhaul traffic due to the WiFi access points. The study explores wide deployment of data cache from massive access points
for serving the several thousand active users. The study reduces the backhaul traffic using deep learning model that conducts statistical
analysis on the collected user records. Extensive simulations are conducted to study the efficacy of the model that includes the
cumulative distribution function per access point traffic/entropy and Jaccard similarity, caching resource utility and cache gain ratio.

Keywords: Deep Learning, Densenets, Big Data Analytics, Wifi Systems

1. Introduction spectral sensors for capturing frame data, the information

accuracy is reduced in several ways. When it comes to
There has been an increase in the creation of mobile data
gathering data from networks, this does not modify the
across wireless networks due to the increasing use of
network state, but it may be used to create context-aware
mobiles. IEEE 802.11 networks are used by a substantial
apps that can help with monitoring at all levels. Passive
number of end-users in a variety of settings [1] [2].
scanning of beacons, together with information about the
Large-scale wireless networks like this one provide
received signal power, allows the identification of
valuable information about users, their networks, their
neighbouring access points and the inference of their
usage, and their mobility patterns. [3]
physical distance and radio coverage area.
The monitoring of a wireless LAN is more difficult than
Use of user positioning or geolocation information helps
that of a wired LAN because of the inherent limitations
to discover usage trends and spot abnormalities, such as,
of wireless monitoring. In the wired world, methods such
for instance, it is also possible to design more efficient
as measuring metrics after data has travelled across the
networks by studying the data produced by the networks
network are tried and proven, but they do not indicate the
themselves [4].
current network state [4].
By analysing a vast detect problems or performance
An active measurement is required in order to accurately
decreases at specific network locations, and optimise
assess a network current condition; this necessitates
channel allocation [5]. An examination of the wireless
changes in the parameters being examined. When using
network environment is a significant challenge because
indirect measures, such as access point counters or
of factors [6]. Ability to extract information about users
1
Assistant Professor, Department of Information Technology, Prasad V.
and networks are hampered by the absence of network
Potluri Siddhartha Institute of Technology, Andhra Pradesh, India. Email
ID: [email protected] intelligence mechanisms.
2
Professor, Department of ECE, Rajalakshmi Engineering College, Tamil
Real-time network analysis necessitates the use of large
Nadu, India. Email ID: [email protected]
3
Associate Professor, Department of ECE, Karunya Institute of data streaming processing techniques, which may not be
Technology and Sciences (Deemed to be University), Tamil Nadu, India. appropriate for other applications. Data quality and their
Email ID: [email protected] statistical distribution are unknown, so this is a problem.
4
Professsor, Department of Computer Science and Engineering, DMI
College of Engineering, Chennai. Email ID: [email protected] A huge data processing platform processes a limited
5
Assistant Professor, Faculty of Computing and Informatics, Jimma dataset at a time in batch processing, which is the
Institute of Technology, Jimma University, Ethiopia. Email ID:
[email protected]
opposite of what is done in stream processing. As a
6
Assistant Professor, Department of ECE, Gokaraju Rangaraju Institute of result, batch processing necessitates a lot of storage
Engineering and Technology, Telangana, India. Email ID: space and has a noticeable impact on reaction time. Each
[email protected] data sample is processed at a shorter latency, and there
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(7s), 326–331 | 326
are no memory limits on the storage when using Wireless networks have recently been studied using DL
streaming data processing. to learn more about their structure and behaviour. In
order to investigate bi-modal data, created an algorithm
The limitless number of samples and the demand for
based on DL. Devices' indoor location fingerprints are
minimal delay while processing fresh samples
generated by their algorithm. Bi-modal data is used to
necessitate that classic machine learning methods be
train DL to develop feature-based fingerprints offline. In
modified for the streaming data. Streaming data leads to
the deep autoencoder network, each point has its own
learning problems in deep learning algorithms. It is
unique fingerprint.
important to keep an eye out for changes in the statistical
data distribution in deep learning applications. In this One solution uses an Auto Encoder, while the other uses
work, we have developed using DenseNets that reduces Convolutional Neural Networks (CNNs) for indoor
the backhaul traffic due to the WiFi access points. localization. The authors conclude that the Auto Encoder
technique has a lower level of error. It employ DL to
The main contribution of the paper is given below:
detect wireless channel activity more accurately and
• The study explores wide deployment of data cache robustly. The key concepts are selecting high-quality Wi-
from massive access points for serving the several Fi channels and switching between them to construct an
thousand active users. extended channel. Using a recursive neural network
model, the authors first look for trends in the usage of
• The study reduces the backhaul traffic using deep
specific channels.
learning model that conducts statistical analysis on
the collected user records. 3. Proposed Method
• Extensive simulations are performed. Numerous computational deep learning models have
been developed in this context to address a variety of
2. Related Works
workloads. Static large data can be processed with batch
User mobility characterization applications are classified data processing techniques. In terms of batch data
by the author of [6], attempting to forecast the person processing, Hadoop is the most commonly used tool.
future location and mobility patterns, only the number of Analyzing network data necessitates the use of systems
people going through a given location is taken into that can analyse data in real time.
account. User mobility patterns are identified by doing a
3.1. DenseNet
spatial-time analysis, which aims to discover the groups
of users with similar characteristics and their future paths It was set to 4 in this case because that what the ResNet-
[7]. In [8] claim that in order to maximise the QoS of 18 model suggested. Each dense block will have 128
mobile users, next-generation communications will more channels as a result of this study convolutional
employ. Because of this, the authors advise that users layer channel count of 32. As a result, the height and
should be able to easily move between different access width are reduced by half, as well as the number of
points or base stations on their mobile devices. Using an channels.
MDP, the proposal selects the best networks to enhance
For functions, the study recalls the Taylor expansion. In
QoS, they utilise an algorithm that uses reinforcement
relation to x = 0, the expression for the point is given
learning to determine a better option.
below:
The Q-learning technique is in the article [9] to propose a
f(x)=f(0)+f′(0)x+f′′(0)2!x2+….
method each user sets a bit in a bit vector, and the state
set for Q-learning spectrum sensing contains all of the The main point is that a function can be decomposed
possible combinations of bits in that vector. The action f(x)=x+g(x).
set for Q-learning is indexed from 0 to 1, with 0 denoting
an available state for secondary users and 1 denoting an A basic linear term is created by ResNet, and a nonlinear
unavailable state. When the action is in line with the term is created by ResNet. Is there a way for the study to
channel occupation, the rewards are positive; otherwise, catch information beyond the two terms? DenseNet was
they are negative. one such solution.

In addition, the Q-Learning method is used to implement Differences are primarily based on the fact that outputs
bandwidth control for cloud providers. For one, they are concatenated rather than appended in DenseNet. As a
regularly adjust network infrastructure parameters to result, after applying a series of more complex functions,
match service level agreements with clients, while for the research performs a mapping to its values. DenseNet
another, they maximise network occupancy and, is the name given to this algorithm because the
consequently, income for the cloud provider. dependence block between variables gets quite thick.

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(7s), 326–331 | 327
The final link in this chain is tightly linked to all the definitions, so the study will use them to examine the
links that came before it. results. Using the Python package, the study were able to
achieve this for each block by calculating the collection
S`
of selected features. The functions were used to compute
Dense Layers: the features in various preprocessing phases. All edges
It uses a modified activation, batch normalisation and having an RSSI value lower than-80 dBm, which is
convolution structure similar to ResNet. Convolution considered the Channel Assessment criterion for APs,
blocks with the same number of output channels make are first discarded from consideration. Edges between all
into a dense block. A convolution block input and output of the nearby APs that do not impact one another are thus
are combined in a forward propagation investigation, erased in this manner. The following features are then
though. calculated:

Transition Layers: Number of edges

Adding too many transition layers will result in an overly Links with RSSI values greater than or equal to 80 dBi
complicated model, since each thick block increases the are counted as having this value. For each pair of APs
number of channels. Transition layers are utilised to keep that can see each other above the threshold, the
the model simple. symmetry should hold.

Training: In-degree centrality

The study will reduce the input 224×96 because it is There was yet another preprocessing step required for the
utilising a deeper network here. remaining features because the edge weights had to seem
like a distance measure in those cases. In order to get the
3.2. Data Analytics and Classification distance between APs, i.e., the adjacency matrix entry,
RSSI readings from several office Wi-Fi networks were the study use Eq.(1) to convert the RSSI values (dBm).
used to generate the dataset. One-to-one computing is This distance is then used in Eq. (1) to get the distance
being implemented in Uruguay by this organisation. So, between APs.
one of its most important duties is to ensure that all wij=0.1(10−RSSIij) (1)
educational institutions in the country have access to Wi-
Fi internet access. Each Wi-Fi network AP conflict block can be interpreted
in many ways, and each feature picked has a possible
Most Wi-Fi networks are found in public primary and interpretation. All nodes have an influence on their
secondary offices, and it is crucial to keep this in mind. neighbors, which is reflected in all centrality
These educational institutions are housed in a wide range measurements, regardless of the disparities between
of buildings, from century-old structures with multiple them. This means that in a Wi-Fi network, an AP with a
levels. Each building is normally covered by 5 or 6 APs high centrality value is more likely to be located in an
on average, but about 20% of the buildings required area with higher levels of interference from its nearby
more than 10 APs, as shown in the graph. APs. When it comes time to deal with key concerns, such
The algorithm is used to manage radio resources by the as channel allocation and transmission power regulation,
WLCs. Radio transmissions of NDP packets use a single it important to know how many channels each AP needs
radio chain with the utmost power and lowest data rate to avoid interference.
permitted for the channel/band. NDP packets are sent As a result, most of the characteristics do not have a high
every 180 s by default on all channels. In the 2.4 GHz connection with each other, preventing the feature vector
range, the AP goes off-channel roughly every 16 seconds from being overly saturated with redundant data. The
to deliver an NDP packet; in the 5 GHz band, it goes off- final step in data exploration and feature analysis is PCA
channel every eight seconds. NDP data and RSSI for all decomposition, which digs deeper into the significance
received packets are forwarded to the WLC. The WLC of each characteristic found in our sample set.
takes five measurements each neighbour and averages
them over a period of 15 minutes. Because it is busy and 4. Results and Discussions
has less available spectrum than the 5 GHz band, the In this section, the study validate the proposed DenseNet
study will focus our attention on 2.4 GHz measurements. based on several network metrics that include cumulative
As a result, in order to build the conflict blocks, the study distribution function per access point traffic/entropy and
treat each AP as a node. As a result, each time stamp Jaccard similarity, caching resource utility and cache
corresponds to a single directed block for each office. gain ratio. The model is simulated in python simulator
This study relies on conflict blocks generated by Wi-Fi with eclipse IDE on a high end computing engine. The
network data, which are defined in terms of their proposed model is compared with conventional deep
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(7s), 326–331 | 328
learning models including Deep Auto Encoders (DAE) and Residual Neural Network (ResNet).

1
DenseNet
0.9 DAE
ResNet
0.8

0.7

0.6
CDF

0.5

0.4

0.3

0.2

0.1
5 10 15 20 25 30 35 40 45 50
Average Traffic Consumed per AP (kB)

Fig 1: CDF vs. Consumed Traffic

For a successful cache deployment, it is important to As seen in Figure 1, the findings of the CCDF in two
know where the most traffic is consumed and the distinct months are very similar since the two curves are
consumed traffic. The study collects all user association so close together. It suggests that if the first deployment
records from WiFi to capture the attribute. Each user method is well-planned, the caching gain can last for a
association records are logged for a week on the long time because the demonstrated consistency of traffic
platform, which can collect raw management data from consumption every day is demonstrated to be consistent.
all APs in real time. Second, the amount of AP traffic consumed is spread out
over a wide range of traffic volumes.

1
ResNet
0.9
DenseNet
0.8 DAE

0.7

0.6
CDF

0.5

0.4

0.3

0.2

0.1

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Jaccard Similarity

Fig 2: CDF vs. Jaccard similarity score

Figure 2 shows a strong Jaccard similarity score between APs, caching services can be provided to a greater
popularity ranking and traffic consumption. According to number of users if the cache deployment method is
the findings, AP popularity is positively connected with intended to cache this traffic as well.
its traffic consumption. Since more traffic is generated at

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(7s), 326–331 | 329
 1
ResNet
0.9
DenseNet
DAE
0.8

0.7

CDF 0.6

0.5

0.4

0.3

0.2

0.1

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
APs

Fig 3: CDF vs. APs

The CDF of APs in buildings and edge nodes is shown in cache resources. A specialised cloud centre is not
Figure 3 of the study. Edge nodes are used instead of necessary because the edge node is tiny enough to only
APs for caching so that all AP-associated clients can have a few APs, for example. Edge node caching is
access and benefit from the same cache resource. It has considered easier to create and maintain than a data
the potential to dramatically increase the utilisation of centre in every building.

1
ResNet
0.9 DenseNet
DAE

0.8
Average Gain Caching Ratio

0.7

0.6

0.5

0.4

0.3

0.2
1000 1500 2000 2500 3000 3500 4000
File Size (KB)

Fig 4: Average Caching Gain Ratio

The caching gain ratio plot is shown in Figure 4, and it several thousand active users. The study reduces the
reveals the following three key findings: When two backhaul traffic using deep learning model that conducts
curves are close together, the suggested DenseNet is able statistical analysis on the collected user records. The
to achieve an almost ideal level of performance. cumulative distribution function per access point
traffic/entropy and Jaccard similarity, caching resource
5. Conclusions
utility and cache gain ratio.
In this paper, DenseNet reduces the backhaul traffic in
Pattern recognition training in the future will use labelled
the WiFi access points and it explores wide data cache
patterns to train algorithms. Aside from the
deployment from massive access points for serving the
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(7s), 326–331 | 330
computational costs and time involved in labelling. To [5] Covert, M. W., Gillies, T. E., Kudo, T., &
further complicate things, labels are affected by the Agmon, E. (2021). A forecast for large-scale,
speed and volume of data. One of the most pressing predictive biology: Lessons from
training-related issues is the issue of overfitting, which meteorology. Cell Systems, 12(6), 488-496.
remains an open question.
[6] Midoglu, C., Kousias, K., Alay, Ö., Lutu, A.,
References Argyriou, A., Riegler, M., & Griwodz, C.
(2021). Large scale speedtest experimentation
[1] Yang, H. H., Xu, C., Wang, X., Feng, D., &
in Mobile Broadband Networks. Computer
Quek, T. Q. (2021). Understanding age of
Networks, 184, 107629.
information in large-scale wireless
networks. IEEE Transactions on Wireless [7] Fan, C., Yan, D., Xiao, F., Li, A., An, J., &
Communications, 20(5), 3196-3210. Kang, X. (2021, February). Advanced data
analytics for enhancing building performances:
[2] Lv, Z., Lou, R., Li, J., Singh, A. K., & Song, H.
From data-driven to big data-driven approaches.
(2021). Big data analytics for 6G-enabled
In Building Simulation (Vol. 14, No. 1, pp. 3-
massive internet of things. IEEE Internet of
24). Tsinghua University Press.
Things Journal, 8(7), 5350-5359.
[8] Asadianfam, S., Shamsi, M., & Kenari, A. R.
[3] Cheng, S., Ma, L., Lu, H., Lei, X., & Shi, Y.
(2021). TVD-MRDL: traffic violation detection
(2021). Evolutionary computation for solving
system using MapReduce-based deep learning
search-based data analytics problems. Artificial
for large-scale data. Multimedia Tools and
Intelligence Review, 54(2), 1321-1348.
Applications, 80(2), 2489-2516.
[4] Shapsough, S., Takrouri, M., Dhaouadi, R., &
[9] Yin, L., Lin, N., & Zhao, Z. (2021). Mining
Zualkernan, I. A. (2021). Using IoT and smart
daily activity chains from large-scale mobile
monitoring devices to optimize the efficiency of
phone location data. Cities, 109, 103013.
large-scale distributed solar farms. Wireless
Networks, 27(6), 4313-4329.

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(7s), 326–331 | 331