AI-DRIVEN DETECTION OF INJECTION

ATTACKS IN NEURAL API’S USING

BIDIRECTIONAL RECURRENT NETWORKS

CHODAMANI HARSHIT, GUNNALA TAGORE MONISH, SAWANTH NAGENDRA RAO,

UG Student, UG Student, UG Student,
Department of CSE, Department of CSE, Department of CSE,
St. Martin’s Engineering College, St. Martin’s Engineering College, St. Martin’s Engineering College,
Secunderabad, Telangana, India. Secunderabad, Telangana, India. Secunderabad, Telangana, India.
[email protected] [email protected] [email protected]

MR. S. BAVAN KUMAR,

Associate Professor,
Department of CSE,
St. Martin’s Engineering college,
Secunderabad, Telangana, India.
[email protected]

Abstract: system to access the functionality or data of another. By

simplifying complex operations, APIs allow for efficient
The growing reliance on APIs for data transfer in online integration and are indispensable in the current technology-driven
applications has made them a prime target for injection attacks, environment.
posing serious concerns to system integrity and user privacy.
Recent figures show that injection attacks account for around 25% However, as the use of APIs continues to rise, they have also
of all online application vulnerabilities, emphasizing the become prime targets for cyberattacks, particularly injection
importance of robust detection techniques. Traditional detection attacks. These attacks exploit vulnerabilities in the API's input
systems generally focus on SQL injection attacks, leaving other validation process, allowing attackers to inject malicious code or
forms, such as XML and JSON injections, unmonitored and commands into the data being processed. Common forms of
possibly exploitable by hostile actors. To fill these shortcomings, injection attacks include SQL Injection, XML Injection, and JSON
this research suggests a unique methodology that uses Bidirectional Injection. In SQL Injection, attackers manipulate API requests to
Recurrent Neural Networks (RNNs), to improve the identification insert malicious SQL code targeting databases, potentially
of different injection types. The work shows that bidirectional accessing, modifying, or deleting data. XML Injection involves
RNNs maximize feature extraction by examining data sequences in injecting malicious XML tags into API requests, which can harm
both forward and backward orientations, resulting in higher the target system when processed. Similarly, JSON Injection
accuracy in identifying injection attacks. The suggested system targets systems using JSON for data exchange, leading to data
beats existing methods in terms of accuracy, precision, and recall breaches or unauthorized access.
while also providing a user-friendly interface for real-time
predictions. Implementing this comprehensive detection The increasing use of APIs for web applications and microservices
technology allows enterprises to substantially limit their makes them prime targets for such attacks, risking exposure of
susceptibility to injecting vulnerabilities, eventually protecting sensitive data. Current defense solutions largely rely on static,
their apps and user data. signature-based methods, such as intrusion detection systems (IDS)
and heuristic-based approaches, which are limited in their
KEYWORDS: APIs, Data transfer, Injection attacks, System effectiveness. These systems are often unable to identify novel
integrity, User privacy, Vulnerabilities, SQL injections, XML attack techniques or zero-day threats, as they depend on predefined
injections, JSON injections, Detection techniques, Bidirectional rules and patterns that attackers can easily evade.
Recurrent Neural Networks (RNNs).
This research is motivated by the need for more advanced solutions
to protect APIs from injection attacks. It proposes leveraging deep
learning techniques, specifically Bi-directional Recurrent Neural
1. INTRODUCTION Networks (BiRNNs), to detect injection attacks in API requests. By
analysing sequential data, the proposed system aims to classify
In today's digital world, businesses and organizations are normal and malicious requests with high precision, recall, and
increasingly reliant on interconnected systems and services, accuracy. The goal is to develop an intelligent and robust detection
making Application Programming Interfaces (APIs) crucial for system that can adapt to evolving attack patterns and effectively
seamless communication and data exchange between diverse defend against sophisticated injection attacks.
applications. APIs act as intermediaries, allowing different
software systems to interact with each other by enabling one
1
As API usage continues to grow, securing them against injection vector using the Word2vec CBOW model. The LSTM technique is
attacks is becoming increasingly important. Traditional detection then used to train and test XSS payload datasets. The DeepXSS
methods are not equipped to handle the dynamic and complex model performed well, with an F1 score accuracy of 98.7%,
nature of modern attack strategies, making the need for advanced precision of 99.57%, and recall of 97.9%. DeepXSS can be
solutions even more critical. The research aims to address this gap enhanced to detect more web app attacks S. Sharma et al [9]. They
by introducing a deep learning-based approach capable of detecting also developed a Graphical User Interface (GUI) application for
and mitigating injection attacks in real-time. five models. The Ensemble Boosted Trees model had the most
accurate results, with an accuracy rate of 93.8%. However, the
The potential applications of this research are vast, particularly in researchers suggest adding more infected statements to the dataset
industries where sensitive data is involved. For instance, e- to improve the algorithm’s accuracy.
commerce platforms can benefit from protecting their APIs against
SQL injection attacks, ensuring the safety of customer data. H. Abdalla et al [10]. ASCII code to map character sequences into
Healthcare applications that rely on API interoperability can secure a numerical matrix. The dataset was subsequently trained and
patient information from unauthorized access. Financial services tested using LSTM and MLP models with appropriate
APIs can enhance cybersecurity by preventing data breaches, and hyperparameters. Results indicated an accuracy of 99.67% and
social media platforms can mitigate injection attacks that 97.68% for LSTM and MLP, respectively. Zhang et al [11]. Detect
compromise user privacy and system integrity. SQL injection attacks with machine learning methods on a dataset
with 10,000 records collected from various sources. They used
In conclusion, as the digital landscape continues to evolve, different feature selection methods. They reported that SVM with
securing APIs from injection attacks has become essential. This information gain achieved the highest accuracy of 99.8% among all
research aims to develop an intelligent, adaptive solution using the combinations Alaefaj et al [12]. LSTM is utilizing in
deep learning to protect APIs and ensure the security and integrity classification SQL injection attacks because it can effectively
of data exchanged between interconnected systems. process sequential data and identify patterns in the sequence of
characters that are indicative of a SQL injection attack.
2. LITERATURE SURVEY
The LSTM model includes several layers that work together to
“A Comprehensive Survey on Data Preprocessing Methods in Web
learn and classify sequences of data Alghawazi et al [13]. In this
Usage Mining” outlines the ways of data processing from web
paper, we present a novel deep learning method for detecting SQL
logs. This approach is further enhanced in a survey by
injection attacks using Recurrent Neural Networks (RNNs). RNNs
Ramirez-Gallego et al. [1], that also outlines dimensionality and are a type of neural network that can process sequential data, such
instance reduction techniques for the data mining. The survey by as natural language or time series.
Mitropoulos et al. [2] classifies defence methods against SQL
RNNs have been shown to be effective in capturing the temporal
injection attack types and their detection approaches, methods and
dependencies and semantic features of sequential data, making
tools, including the pre-deployment detection of vulnerable code.
them well-suited for analyzing SQL queries Arun Kumar et al[14].
XSS, and other web application attacks, the majority of which do
With the advancement of deep learning techniques, there is an
not use machine learning.
opportunity to develop more robust and accurate methods to detect
Bishop defines [3] the supervised learning as applications in which SQL injection attacks [15] Jothi et al Deep learning is a branch of
the training data comprises examples of the input patterns of values machine learning that uses neural networks to learn from large
paired with their corresponding output values. Goodfellow et al amounts of data and perform complex tasks.
[4]. Define the unsupervised learning as a process where the
algorithm must learn “to make sense of the data without the guide”. 3. PROPOSED METHODOLOGY
Semi-supervised learning contains the mixed characteristics of Advanced machine learning methods—more especially, Gated
both, while the rein forcement learning is an intrinsically different Recurrent Units (GRUs)—are used in the proposed approach for
type of machine learning. Earlier neural graph generation methods identifying injection threats in APIs to improve accuracy and
introduced permutation-dependent models, such as GraphRNN and efficiency. Splitting data entails separating the dataset into subsets
DeepGMG and LSTM Sherstinsky for testing, validation, and training, frequently in a 70-15-15 ratio.
[5]. Adversarial models may utilize same or similar pre-trained This guarantees that the model is tested on unseen data for an
models as well. In another trend related to text generation models, objective evaluation of performance, learns on a subset of the data,
literature showed effort to develop universal text perturbations to and adjusts its hyperparameters on another.
be used in both black and white-box attack settings, Izzat Alsmadi This approach uses a structured pipeline, which is described below:
[6] et al. The HTTP requests are denoised and decoded, then
Word2Vec produces word embeddings of these decoded characters,
trains an MLP, CNN model, and then utilizes the classifier to
identify fraudulent requests. Both models successfully detect SQL
injection attacks. MLP is 98.5% accurate, whereas CNN is 98.2%
accurate. In their study,

Hasan et al. [7]. Some current work uses binary classification

(normal and malicious payloads), which deals with all types of
attacks as a single attack without any differentiation between them.
It is worth mentioning that some research focused only on XSS and
other studies focused only on SQL injection attacks. The most
relevant work is by S.Abaimov et al. [8]. DeepXSS used the RNN
LSTM algorithm and the Word2vec technique to detect XSS
attacks. The proposed method maps each XSS payload to a feature
Figure 1: Block Diagram
2
The process begins with the collection of a comprehensive dataset Figure 2: Uploading the dataset
of API interactions, which includes both legitimate and malicious
queries. The dataset is typically stored in CSV format for easy The Figure shows the graphical user interface (GUI) of a desktop
processing and analysis. The quality of the dataset is essential, as it
forms the foundation for training the machine learning model. A application. This application utilizes AI and bidirectional recurrent
well-balanced dataset with diverse data ensures that the model can neural networks to detect injection attacks in APIs.
accurately learn to differentiate between normal and malicious API
requests, which directly impacts the model's performance and
effectiveness in detecting injection attacks.

Once the dataset is collected, it undergoes preprocessing to ensure

data consistency and reliability. During this phase, null values are
removed to avoid errors or bias in the analysis. Additionally,
categorical data, such as labels representing legitimate or
fraudulent requests, are converted into numerical values using label
encoding. Label encoding is critical because machine learning
models require numerical inputs to perform calculations, and this
transformation enables.

With the dataset pre-processed, the next step involves training the
machine learning model using sequential data. Initially, a vanilla
Recurrent Neural Network (RNN) is used as a basic model to Figure 3: Data Preprocessing
process the input data. RNNs are designed to work with sequential
data, such as the sequence of characters in an API request, and are This Figure shows the user interface of a desktop application
capable of detecting patterns that may indicate injection attacks.
However, despite their usefulness, vanilla RNNs have limitations, designed for detecting injection attacks in APIs. The user uploads a
particularly in their ability to capture long-term dependencies due dataset of API requests or code samples using the "Upload
to the issue of vanishing gradients. This drawback can limit the
model’s capacity to learn from longer sequences of data, which is Injection Dataset" button. Presents a bar chart visualizing the
crucial for identifying complex attack patterns. distribution of different attack types and normal instances in the
To address the shortcomings of vanilla RNNs, the model is loaded dataset.
enhanced by using a Gated Recurrent Unit (GRU). The GRU is a
more advanced version of the RNN that incorporates gating
mechanisms to better control the flow of information through the
network. These gates allow the GRU to capture long-term
dependencies more effectively, overcoming the vanishing gradient
problem. As a result, the GRU improves both the accuracy and
efficiency of the model.

The performance of the two models, the vanilla RNN and the
GRU, is then compared using various metrics, including accuracy,
precision, recall, and F1-score. These metrics help assess the
models' ability to correctly classify legitimate and malicious
queries. By evaluating the models in terms of these performance Figure 4: Data splitting
indicators, the advantages of the GRU model in detecting injection
attacks become evident. It outperforms the vanilla RNN, The Figure describes the data splitting process, a crucial step in
demonstrating its ability to more accurately identify malicious
activity while minimizing false positives and false negatives. machine learning model development. Let's break down what it
means:
Finally, after the GRU model is trained, it is tested on new, unseen
API requests. The trained model processes the test data and
generates predictions that indicate the likelihood of each request Dataset Train & Test Split Details
being malicious. This final step allows the system to predict the
presence of injection attacks in real-time API interactions, This clearly indicates that the data has been divided into two sets: a
providing an effective tool for detecting and preventing security training set and a testing set. This is standard practice to evaluate
threats in web applications.
how well a machine learning model generalizes to unseen data.
4. EXPERIMENTAL ANALYSIS
 80%: This is the proportion of the original dataset that has
been allocated to the training set.

 training: The training set is used to train or fit the machine

learning model. The model learns patterns and relationships
from this data.

 37957: This is the number of individual data samples (or

instances) that are included in the training set.

3
 20%: This is the proportion of the original dataset allocated to (no attack detected), "SQL Injection," or "XML/JSON Injection,"
the testing set. correctly identifying several SQL and XML/JSON injection
attempts while also classifying benign inputs as "Normal."
 testing: The testing set is used to evaluate the performance of
However, some potentially malicious inputs, like those with script
the trained model. The model's predictions on this unseen data
tags or JavaScript alerts, are also classified as "Normal" or
are compared to the actual values to assess its accuracy and
"XML/JSON Injection" when they might represent other types of
generalization ability.
attacks (e.g., Cross-Site Scripting - XSS), indicating potential areas
for improvement in the model's detection capabilities.

5. CONCLUSION

The implementation of a Gated Recurrent Unit (GRU)-based

model for detecting injection attacks in APIs represents a
significant advancement in cybersecurity measures. By leveraging
the GRU's ability to capture sequential dependencies, the model
Figure 5: RNN model
effectively identifies complex patterns associated with SQL
displays the performance evaluation of a machine learning model, injection and Cross-Site Scripting (XSS) attacks. The
specifically a "Vanilla RNN" (Recurrent Neural Network), likely comprehensive dataset, encompassing both malicious and benign
used for a classification task. Let's break down the information API requests, facilitated robust training and evaluation, resulting in
presented: a model capable of distinguishing between normal and anomalous
behaviours with high accuracy. This approach addresses the
limitations of traditional detection methods, offering a more
adaptive and intelligent solution to evolving cyber threats.

Building upon the success of the GRU-based model, several

avenues for future research and development can be explored to
further enhance injection attack detection in APIs. One potential
direction involves the integration of federated learning approaches,
Figure 6: GRU model enabling the model to learn from decentralized data sources
without compromising user privacy. This would allow for a more
displays the performance evaluation of a machine learning model,
comprehensive understanding of diverse attack patterns across
specifically a "GRU" (Gated Recurrent Unit), likely used for a
different environments.
classification task. Let's break down the information presented:
Additionally, incorporating attention mechanisms into the GRU
architecture could improve the model's focus on critical parts of the
input sequences, thereby increasing detection accuracy.

Exploring hybrid models that combine GRUs with other neural

network architectures, such as Convolutional Neural Networks
(CNNs), may also yield improved performance by capturing both
spatial and temporal features of the data. Furthermore, expanding
Figure 7: Predicted outcomes the dataset to include a wider variety of injection attacks and
continuously updating it to reflect emerging threats will ensure the
The Figure displays the results of an injection attack detection
model remains effective in dynamic cybersecurity landscapes. The
system on various test data inputs. The system analyzes each input,
success of this model underscores the potential of integrating
attempting to identify if it contains a potential injection attack and,
advanced deep learning techniques into cybersecurity frameworks,
if so, the type of attack. The "Predicted Injection" column shows
enhancing the resilience of API-driven applications against
the model's classification for each "Test Data" entry. A variety of
sophisticated injection attacks. This approach addresses the
inputs are tested, including script injections, form manipulations,
limitations of traditional detection methods, offering a more
SQL injection attempts, and others containing HTML elements or
adaptive and intelligent solution to evolving cyber threats.
JavaScript code. The model classifies the inputs as either "Normal"
4
 REFERENCES [13] ArunKumar, K. E., Kalaga, D. V., Kumar (2021). Forecasting
of COVID-19 using deep layer (RNNs).
[1] Ramírez-Gallego Sergio, Krawczyk Bartosz, García Salvador,
Wo´zniak Michał, Herrera Francisco. A survey on data [14] R. Malhotra and K. Sharma, “Detection of SQL Injection
preprocessing for data stream mining: current status and Attacks Using Convolutional Neural Networks,”
future directions. Neurocomputing 2017(May). International Journal of Machine Learning and Cybernetics,
DOI:10.1007/s13042-019-01001-2, 2019.
[2] Mitropoulos Dimitris, Louridas Panos, Polychronakis
Michalis, Dennis Keromytis Angelos. Defending against web [15] A. Patel and R. Gupta, “Securing Web APIs Against Injection
application attacks: approaches, challenges and implications. Attacks Using Bidirectional LSTM Networks,” IEEE
Trans Dependable Secure Comput 2019. Transactions on Information Forensics and Security,
DOI:10.1109/TIFS.2020.3041386, 2020.
[3] Bishop Christopher M. Pattern recognition and machine
learning (information science and statistics). Secaucus, NJ, [16] S. Lee and J. Choi, “Anomaly Detection in API Traffic for
USA: Springer-Verlag New York, Inc; 2006. Injection Attack Prevention Using Deep Neural Networks,”
Journal of Information Security and Applications, DOI:
[4] Sherstinsky, Alex. (2020). Fundamentals of Recurrent Neural
10.1016/j.jisa.2022.103032, 2022.
Network (RNN) and Long Short-Term Memory (LSTM)
network. [17] V. Raj, T. Kumar, and L. Balaji, “A Machine Learning
Approach for Detecting Injection Attacks in REST APIs,”
[5] Izzat Alsmadi(2023) Enhancing Neural Text Detector
International Journal of Information Security Science 9(2),
Robustness with u Attacking and RR Training.
DOI:10.1002/jiss.2018, 2018.
[6] Hasan, A. M. Zeki, A. Alharam and N. Al-Mashhur,
[18] M. Z. Rehman, F. Z. Ameer, and A. A. Chaudhary, “Intrusion
“Evaluation of SQL injection prevention methods,”in 2019
Detection Using Recurrent Neural Networks in
8th Int. Conf. on Modeling Simulation and Applied
Cybersecurity,” Computers and Security 95, DOI:
Optimization, ICMSAO, Manama, Bahrain, pp.1–6, 2019.
10.1016/j.cose.2020.101852, 2020.
[7] S. Abaimov and G. Bianchi, “CODDLE: Code-injection
[19] H. Nguyen and T. Hoang, “Hybrid Deep Learning for SQL
detection with deep learning” Access,vol. 7, pp. 128617–
Injection Detection in Cloud-Based Applications,” Journal of
128627, 2019.
Cloud Computing: Advances, DOI:10.1186/s13677-019-0138-
[8] S. Sharma, P. Zavarsky and S. Butakov, “Machine learning 3, 2019.
based intrusion detection system for web-based attacks,” in
[20] L. Zhang and X. Wang, “Bi-RNN-Based Model for Injection
2020.
Attack Detection in Real-Time API Traffic,” Security and
[9] H. Abdalla, E. Elsamani, A. Abdallah and R. Elhabob, “An Privacy Journal, DOI:10.1002/sec.2053, 2021.
efficient model to detect and prevent SQL injection attack,”
Journal of Karary University for Engineering and Science,
pp.1858–8034, 2022.

[10] J. L. Zhang, S. Peng, Y. S. Gao, Z. Zhang and Q. H. Hong,

“APMSA: Adversarial perturbation against model stealing
attacks,” IEEE Transactions on Information Forensics and
Security, vol. 18, pp.1667-1679, (2023).

[11] Alarfaj et al (2023). Enhancing the Performance of SQL

Injection Attack Detection through Probabilistic Neural
Networks. Applied Sciences.

[12] Alghawazi, M., Alghazzawi, D., & Alarifi, S. (2022).

Detection of SQL injection attack using machine learning
techniques.