NTU Journal of Engineering and Technology

Structured Query Language Injection (SQLi) remains one of the most
serious threats to web applications and has the ability to bypass traditional
signature-based detection through obfuscation and zero-day payloads. This has driven the wider application of Machine Learning (ML) and Deep
Learning (DL) techniques. This paper analyzes 50 peer-reviewed
literatures published in the interval between 2015 and 2025, where the
reported accuracy of detection ranged between 93 and 99.9%. Traditional
ML methods include Support Vector Machine (SVM), Random Forest
(RF), Logistic Regression (LR), and Decision Tree (DT). DL approaches
encompass Convolutional Neural Networks (CNN), Long Short-Term
Memory (LSTM), Bidirectional LSTM (BiLSTM), and transformer-based
models such as Bidirectional Encoder Representations from Transformers
(BERT). Feature extraction methods include Term Frequency-Inverse
Document Frequency (TF-IDF), Word2Vec, and contextual embeddings.
Evaluation of proposed models uncover new research opportunities in
terms of lack of data availability, the problem of calss imbalance, real-time
application, and excessive use of hardware resources.

This article reviews several parts of previous studies, including both
numerical and experimental investigations that examined the effects of
heat transfer and pressure drop when using nanofluids or conventional
fluids. The study presents design parameters, practical scales, and
summarized results of earlier work in tabular form, accompanied by
appropriate discussion. It also examines the influence of different types of
nanoparticles(including metal oxides, nonmetals, and metals)on the basic
thermodynamic and hydraulic behavior of working fluids, as well as the
role of base fluids such as water and air. The review highlights that some
research topics remain attractive and worth further exploration due to their
simple design and high performance. In contrast, other topics are hindered
by limitations, such as the challenges of applying complex geometries or
the negative effects of increased pressure drop and nanoparticle
agglomeration on channel walls, which can obstruct heat transfer. Overall,
this review provides a foundation for more advanced studies on the
performance of nanofluids in corrugated channel systems. These insights
are valuable for guiding the design of compact and energy-efficient heat
exchangers, with potential benefits in reducing energy consumption and
environmental impact. Future research should focus on improving
nanoparticle dispersion techniques, mitigating pressure drop penalties, and exploring hybrid nanofluids and innovative channel geometries to achieve even better hydrothermal performance.

This work proposes an efficient handwriting-person-based identification
scheme amalgamating deep learning and regular machine learning
classifiers. 6,955 fine-quality Arabic handwriting samples were gathered
from 107 users. Deep features were extracted using Convolutional Neural
Networks (CNNs), and SVM, RF, and Logistic Regression (LR) were
deployed for classification. The test accuracy of feature extraction with
CNN was 99.92%, while CNN+SVM reported a maximum
classification accuracy of 99.95%, CNN+LR reported 98.90%, and
CNN+RF exhibited 98.11%. The method has also been implemented as
GUI program to facilitate user-uploaded writings-based real-time
identification. Evaluation parameters such as accuracy, precision, recall,
and F1-score confirm the effectiveness and robustness of the model. The
findings confirm the potential of handwriting-based biometrics for secure
identification and provide future directions for expanding to multilingual
datasets and large-scale deployment.

Electroretinography (ERG) is an essential tool for assessing retinal
function, with responses from photoreceptors, ganglion cells, and inner
layers. Clinical applications are often secondary to structural imaging,
though dysfunction may appear before anatomical changes. This review
compares three ERG types: full-field (ffERG), patterned (PERG), and
multifocal (mfERG), highlighting differences in response, waveform
components, and clinical uses. This review analyzes more than 60 studies
(2014–2025). Advanced analyses in the time, frequency, and time
frequency domains demonstrated diagnostic accuracies between 85% and
97% for early detection of retinal dysfunctions such as glaucoma and
retinitis pigmentosa. Integrating ERG with Optical Coherence
Tomography (OCT) improved structure–function correlation by 15–25%.
The findings highlight that combining ERG with quantitative feature
extraction and OCT enhances early diagnosis and monitoring of retinal
diseases and supports standardized clinical applications.

Orthogonal Frequency Division Multiplexing with Index Modulation
(OFDM-IM) is an effective communication technology that utilizes
subcarrier indices to transmit information in addition to the information
traditionally modulated on active subcarriers. However, OFDM-IM
suffers from high peak-to-average power ratio (PAPR), which can result
in signal distortion and hence system performance degradation. One of the
techniques used to alleviate the PAPR in OFDM-IM is the selected
mapping (SLM) technique. The effectiveness of the SLM PAPR reduction
technique is affected by OFDM-IM system parameters, such as the number of subcarriers and subcarrier activation ratio. This paper explores the impact of these OFDM-IM system parameters on the PAPR reduction
performance of the SLM technique with different phase sequence settings.
The obtained results provide an overview of the collective effect of both
OFDM-IM and SLM parameters in different operation cases. The findings
show that the amount of PAPR reduction is mainly affected by the number
of subcarriers, and less affected by the subcarrier activation ratio. That is
a greater PAPR reduction of up to 4.2 dB is achievable with relatively
smaller number of subcarriers. The results show the limited effect of more
complex phase vectors, hence suggesting simpler SLM parameter setting
alternatives to achieve a target PAPR reduction performance. Finally, the
results show the improvement in Bit Error Rate (BER) performance of
OFDM-IM system employing the SLM with carefully selected PAPR
reduction parameters.

Internet of Things (IoT) connections depend on the Message Queuing
Telemetry Transport (MQTT) protocol; however, it might be difficult to
determine the best Quality of Service (QoS) level in dynamic network
contexts. Convolutional Neural Networks (CNN) and Long Short-Term
Memory (LSTM) networks are combined in this study's adaptive deep
learning framework to optimize MQTT QoS in real time. To depict a
variety of IoT circumstances, such as resource limitations, high load,
network instability, and regular operations, we created a thorough labeled
dataset of 50,000 synthetic samples. In all investigated scenarios, the
hybrid CNN-LSTM architecture maintained 71.44% resource efficiency
while achieving 92.7% validation accuracy in QoS prediction. Our system
showed great confidence in adapting to essential applications (98.83%),
low-resource environments (99.78%), and high-load conditions (99.99%).
For industrial IoT deployments in smart manufacturing, healthcare
monitoring systems, and critical infrastructure management, where
dependable communication under fluctuating resource constraints is
crucial for operational efficiency and safety, this adaptive QoS
optimization framework shows great promise. The suggested architecture
greatly improves network performance while preserving reliability by
providing a workable solution for autonomous QoS control in resource
constrained IoT installations.

In modern vehicle automatic systems, noise interference presents a significant obstacle to the precision and dependability of sensor-based control and communication. This study offers a comparative performance
evaluation of three adaptive filtering algorithms Least Mean Squares (LMS), Normalized LMS (NLMS), and Recursive Least Squares (RLS), to utilized for adaptive noise cancellation (ANC) under mixed noise conditions. A MATLAB-based graphical user interface (GUI) simulation was created to estimate and illustrate the performance of each method across three noise types: Gaussian, Impulsive and Mixed. The results informed that RLS attained the greatest signal to noise ratio (SNR) enhancement (15.6 dB) and with the minimal mean square error (MSE),
whereas the NLMS offered a proficient equilibrium between velocity and
computing complexity. This research evaluation that appropriateness of
NLMS in real time vehicular control applications and RLS requiring high
accuracy.