Malware represents a real threat to information systems, because of the dam age it causes. This threat is growing today, as these programs take on more complex forms. This means they escape traditional malware detection methods. Hence the need for artificial intelligence, more specifically Deep Learning, which could detect malware more effectively. In this article, we’ve proposed a model for malware detection using artificial neural networks. Our approach used data from the characteristics of machines, particularly computers, to train our Deep Learning algorithm. This model demonstrated an accuracy of around 83% in predicting the presence of malware on a machine. Thus, the use of artificial neural networks for malware detection has shown his ability to assimilate complex, non-linear patterns from data.
With the rapid development of the Internet, malware has become one of today’s major cyberthreats. As the diversity of malware increases, antivirus software is no longer able to meet security needs. As a result, millions of computers are under attack. According to Kaspersky Labs (2016), 6,563,145 different hosts were attacked, and 4,000,000 unique malicious objects were detected in 2015. While traditional methods, such as signature-based static analysis, have been foundational in combating known threats, the dynamic and sophisticated nature of modern malware necessitates more advanced detection strategies. This paradigm shift towards deep learning in malware detection has led to substantial improvements in identifying previously unseen and polymorphic malware variants. However, despite the progress made, there exists a research gap in understanding the robustness, scalability, and generalizability of deep learning models across diverse malware landscapes. By focusing on this crucial research gap, the proposed study seeks to contribute valuable knowledge to the existing literature, providing a nuanced understanding of the capabilities and challenges associated with deploying deep learning for malware detection. Through empirical evaluations and rigorous experimentation, the research endeavors to offer practical recommendations for optimizing deep learning models, thereby fostering the development of more robust and resilient malware detection systems in the face of an ever-evolving cyber threat landscape. This paper presents elements of solutions for detecting malware based on artificial intelligence, mainly Deep Learning.
Malware is a generic term for any type of computer program or software known as “malicious software”. It covers a range of techniques and tools designed to infect and damage their victims’ hardware, silently exploit their resources, extort money and/or steal sensitive and important data. They are used by cybercriminals for financial gain. There are different types of malwares, the best known of which are: Ransomware (Bounaamane, B., & Drissi, Z. (2023)) [
All malware detection techniques can be divided into two categories: static methods based on signatures; dynamic methods based on behavior [
Signature-based analysis is a static method based on predefined signatures. These can be file fingerprints, such as MD5 or SHA1 hashes, static strings, or file metadata. The detection scenario, in this case, would be as follows: when a file arrives on the system, it is statically analyzed by the antivirus software. If one of the signatures matches, an alert is triggered, indicating that the file is suspicious [
Today’s attackers are developing malware that can modify its signature. This characteristic of malware is called polymorphism. Antivirus vendors have therefore had to find another means of detection, based on behavior or heuristics. In this method, the actual behavior of malware is observed during execution, looking for signs of malicious behavior: modification of host files, registry keys, establishment of suspicious connections [
The daily impact of malware is disabling your computers and technological tools; Data loss and theft; Financial losses. Some signs of malware infection are computer slowness; problems shutting down or starting up your computer (slow startup); Infection warnings, often accompanied by purchase requests to remedy the situation; presence of unknown programs on your computer.
The two previous methods are increasingly outdated by the complexity of malware, and a solution based on artificial intelligence has been considered by several studies.
Recent work by Ijaz, Durad & Ismail in 2019 focuses on modern malware, which is represented by its runtime behaviors and collected using the Cuckoo Sandbox tool. The authors demonstrate that better predictive accuracy is achieved with static analysis, due to the predominance of simple malware. However, they point out that dynamic analysis is preferable due to the protection mechanisms used by malware [
Kolosnjaji, Zarras, Webster & Eckert in 2016 applied neural networks to old malware in PE format (32-bit Windows executables). Although the results obtained were interesting, this methodology cannot be implemented in real time due to the need for post-processing to remove repeated actions [
We used Microsoft Malware Prediction dataset, from Kaggle database, to build our prediction models. “Microsoft Malware Prediction” is a set of binary classification data (HasDetections “0” or “1”), containing around 16,774,736 machines and 81 attributes (machine and target identifier “HasDetections”), including 52 categorical attributes (23 attributes are attributes are numerically
coded to protect data confidentiality). Our data source consists of 2 csv files, one of which is the training file (train.csv) and the other the test file (test.csv). Together, these files contain 83 columns or attributes describing various information concerning the hardware and software of each machine. Of these, 53 are numerical values. The train.csv file contains 567,730 machines and 83 attributes, while the test.csv contains 243,313 machines with 82 columns. The overview of our dataset is illustrated by the two images (see
The dataset used is made up mainly of healthy machines, as shown by the distribution of data according to labels, as shown on
The first step consisted of dealing with missing values and outliers, the analysis of the dataset has revealed that more than 44 columns have missing values, and between them around 7 columns have more than 50% missing data. The visual representation below shows how missing values are distributed before data processing (see
The second step consisted of data balancing according to the targeted column. Analyzing the target column, we observe an imbalance of occurrences, namely: 482,571 machines not infected by malware; 85,159 malware-infected machines.
For better learning, we’re going to balance this out by putting them at the same level, as shown in the figure below. At the end of this resampling, we obtain a data set containing as many healthy machines as infected machines (see
The following
The last step is encoding. We used the following policy: for any categorical variable with more than 2 unique categories, we used label encoding. As shown in the illustration of
The model was deployed in a python environment, on a HP Gamer Omen RTX computer, 64 GB Ram, Core i9, 2 Tera SSD, with the following tools: Python 3.9; Jupyter Notebook; Google Colab; These are the following libraries: Matplotlib; Numpy; Pandas; Seaborn; Scikit-learn, TensorFlow, Keras. We used 70% of our dataset to train our neural network, while we used 30% of our data for testing.
To implement our artificial neural network (ANN), we went through several steps: data pre-processing, data scaling and finally data partitioning into training and test samples. The data partitioning is illustrated in the figure below (see
Our ANN, based on a biological model, is represented as follows, with each ball representing a neuron (see
- The set {w0, w1, …, wn} represents the weights associated with each neuron
- The set {x0, x1, …, xn} represents the input values for each neuron
- f represents the activation function
- b represents the neuron’s bias
- The output s of a neuron in each layer is calculated by applying the activation function f to the sum of the input values xk weighted by the weights wk and summed with the bias b.
The model used for our study is an artificial neural network (ANN) whose modeling approach is described in the diagram of
To evaluate the performance of our model, we relied on certain classification evaluation measures, namely:
● The Matrix confusion [
- TP: Our model predicts 84,541 true positives
- TN: Our model predicts that 29,524 machines are not infected, but are in fact infected
- FP: Our model predicts that 105,331 machines will be infected
- FN: Our model predicts that 9248 machines are infected, but they are not.
- Precision score
We trained an Artificial Neural Network (ANN) using Keras to predict the state of machines in relation to malware. Our model achieved an accuracy rate of 83.04%, as shown in
The score was obtained thanks to good training of the neural network. This training gave the following performance (see
On the left-hand side, the error rate decreases, while the accuracy increases over time. The error rate has fallen from around 60% to 20%, demonstrating the accuracy of the model.
● The classification report:
- Precision: Precision is the proportion of instances correctly classified for a given class among all instances predicted to belong to that class.
- Recall: Recall, also known as sensitivity, measures the proportion of instances correctly classified for a given class among all instances belonging to that class.
- F1-score: F1-score is a metric that combines both precision and recall into a single value to assess prediction accuracy. It is particularly useful when classes are unbalanced.
- Support: Support represents the number of real instances for each class, allowing us to evaluate the distribution of the data. We can see all these in
The solution we have proposed for detecting malware within machines or computers enables us to identify the presence of malware. However, the validity of our study may vary depending on whether it is applied to machines equipped with non-proprietary operating systems, such as open-source ones. Indeed, our research has mainly focused on machines running the Windows operating system. Our approach emphasizes machine characteristics, whereas existing studies focus more on the functional aspects of malware. Consequently, extending our findings to open-source environments would require in-depth analysis to assess the feasibility of applying our method.
Given the impact that malware can have on information systems, malware detection is all the more important in cybersecurity.
Z. Zhang, in an article published in 2022, used LightGBM, a machine learning model. His work was based on the same dataset as ours. LightGBM is a boosting model based on the gradient boosting algorithm, designed to improve performance and efficiency over other boosting methods such as XGBoost or Gradient Boosting Machine (GBM). Compared with our work, it scored below ours, which is 0.684 [
D. Huo and colleagues have also carried out research into software detection. They used two learning methods, namely LightGBM and a one-dimensional CNN, with respective scores of 67.16% and 72.47% [
| Model | Score (%) | Authors |
|---|---|---|
| LightGBM | 68.40 | Z. Zhang |
| CNN unidimensionnel 1D | 72.47 | D. Huo, X. Li, L. Li, Y. Gao, X. Li and J. Yuan |
| LightGBM* | 67.16 | D. Huo, X. Li, L. Li, Y. Gao, X. Li and J. Yuan |
| Our ANN | 83.00 |
In this work, we have proposed and implemented a model for detecting malware within machines. This is a deep learning model based on artificial neural networks, to provide a highly efficient model. This work has enabled us to develop a malware detection model using data derived from machine characteristics. The python programming language and TensorFlow library were used to implement this model. The resulting model, based on neural networks, can predict the presence or absence of malware with an accuracy of 83%. The use of this model will enabled malware to be detected effectively, even if it could modify code to evade traditional detection methods and avoid attacks at an early stage.
We would like to thank the Université Virtuelle de Côte d’Ivoire for allowing us to carry out this research work.
The authors declare no conflicts of interest regarding the publication of this paper.
Thiziers, A.H., Tiémoman, K., Gérard, N.B. and Kabir, T.T.Q. (2023) Malware Detection Using Deep Learning. Open Journal of Applied Sciences, 13, 2480-2491. https://doi.org/10.4236/ojapps.2023.1312193