An Efficient Deep Learning Framework for Optimized Event Forecasting

Abstract

There have been several catastrophic events that have impacted multiple economies and resulted in thousands of fatalities, and violence has generated a severe political and financial crisis. Multiple studies have been centered around the artificial intelligence (AI) and machine learning (ML) approaches that are most widely used in practice to detect or forecast violent activities. However, machine learning algorithms become less accurate in identifying and forecasting violent activity as data volume and complexity increase. For the prediction of future events, we propose a hybrid deep learning (DL)-based model that is composed of a convolutional neural network (CNN), long short-term memory (LSTM), and an attention layer to learn temporal features from the benchmark the Global Terrorism Database (GTD). The GTD is an internationally recognized database that includes around 190,000 violent events and occurrences worldwide from 1970 to 2020. We took into account two factors for this experimental work: the type of event and the type of object used. The LSTM model takes these complex feature extractions from the CNN first to determine the chronological link between data points, whereas the attention model is used for the time series prediction of an event. The results show that the proposed model achieved good accuracies for both cases—type of event and type of object—compared to benchmark studies using the same dataset (98.1% and 97.6%, respectively).

1. Introduction

Over the past few decades, terrorism has become a growing problem, resulting in significant destruction that has harmed people’s lives, the safety of their properties, their emotional stability, and the economy [1]. There are several particular features of terrorism, such as the rise of globalization attack expansion and diversification, and many different kinds of targets. Therefore, it is crucial to identify the behavioral characteristics of terrorist groups, investigate global terrorism trends from government data, and enhance emergency management and counterterrorism expertise [2].

The officials of the United Nations defined the word terrorism as “intended to cause physical harm, cause death, or non-military persons to intimidate the public or force the government or international organization to commit or refrain from committing any act” [3]. Terrorism violates law enforcement and creates chaos in affected or targeted regions.

According to Wang et al. [2], terrorist attacks are driven by circumstances related to economics, religion, politics, or society. As reported by the same author, 2111 attacks in total were identified in 2007, which is roughly identical to the all-time high in 1992 [4]. To predict such incidents, artificial intelligence must be utilized to be able to visualize and forecast these kinds of activities because it is one of the youngest disciplines responsible for developing intelligent, customizable algorithms [5].

After the end of the Cold War, terrorist attacks became more widespread. The most significant security threat currently facing the world is the proliferation of many different terrorist attacks, which represents a big cause for concern for humanity as a whole [6].

Particularly, several severe and deadly terrorist attacks have taken place in the past few years, including the terrorist attack on the Kunming Railway Station in China on 1 March 2014 [7]; the bombings of Thailand’s four-faced Buddha on 17 August 2015 [8]; the terror attacks in Paris, France, in November 2015 [9]; the terrorist attack in Nice, France, on 14 July 2019 [10]; and the bombings in Sri Lanka on 21 April 2021 [11]. All of these incidents resulted in significant overall destruction by terrorist attacks.

In recent decades, many studies have investigated terrorism from numerous perspectives, with various techniques [1,2,3,4,9,10,12,13,14,15,16,17,18,19,20] and based on various factors mentioned in the benchmark dataset (details are mentioned in Section 3). The Global Terrorism Database (GTD) [16] has recorded approximately 200,000 terrorist incidents globally, with more than 3000 terrorist organizations attacking 205 nations at multiple levels.

Numerous studies, as mentioned earlier, have focused on terrorism attack prediction. A hidden Markov model was created by (Petroff, et al.) [17] for terrorist attack alerts. Using GTD data from 1970 to 2014, (Meng, et al.) [18] suggested an optimized hybrid classifier method to predict the attack mode of terrorist acts. Data collection, pre-processing, hybrid classification, mining, and classifier testing were all included in the framework. An RP-GA-XGBoost method was suggested by [13] to determine whether innocent bystanders would perish as a result of a terrorist attack. A novel approach for locating primary terrorist organizers was provided by the study. They collected data from many sources of actual terrorist attack incidents to map details of terrorist activity. To gain insights into these networks, they gathered useful information from sources that were accessible and then mapped it as terrorist networks [19].

Lu, et al. [20] applied 3D agents to disclose the actual conduct of people, such as concealing, fighting, and escaping throughout every step against terrorism, using the Peshawar shooting in 2014 as an example. The best possible result from the simulations provided strong evidence for the validity and resilience of the model. They employed counterfactual studies and obtained results showing the precise effects of important factors or mechanisms.

A few researchers have focused on determining the danger and seriousness of harm. Using a stepwise regression methodology and taking into account the effects of vital infrastructure and population concentration, Chatterjee et al. [21] established a regional risk model. A Bayesian network was suggested by Zhu et al. [15] to conduct risk assessments of chemical terrorism strikes. Using scenario analysis and sensitivity analysis, the model was verified, and the impact of important elements was examined. Al-Dahash et al. [22] investigated how communities perceived risk in the wake of terrorist strikes in Iraq. A model to estimate deaths from terrorist attacks was published by Kebir et al. [23].

The challenges of data collection in terrorism research have made open-source datasets particularly valuable [24]. A novel approach to learning, referred to as “Deep Learning” (DL), has emerged from artificial intelligence (AI) and machine learning (ML) [25]. A form of machine learning called deep learning (DL) uses numerous hidden layers to identify intricate patterns in vast amounts of data. Depending on the nature of the challenge, DL can be employed as either supervised or unsupervised learning. ML approaches have already been applied to some aspects of terrorism [26]. The goal is to study and predict future terrorist attacks by using AI techniques and categorization models [27].

To forecast potential acts of terrorism, a hybrid DL model is proposed in this paper. The available work is restricted due to its inability to date to integrate diverse factors, such as weapon type and attack type, in forecasting terrorist activities. The outcomes are subject to terrorist approximations, sample choice, and the statistical method employed. As a result, the current model in [4] can merely be used to restrict the real-world formulation of counterterrorism strategies because it is unable to predict terrorist incidents with an acceptable level of accuracy. Previous studies forecasted terrorist activity using a hybrid DL model. To solve this problem, we suggest a hybrid model that predicts terrorist attacks by combining deep learning, a convolutional neural network (CNN), long short-term memory (LSTM), and an attention mechanism. Additionally, we suggest an enhanced hybrid DL model as found that while the LSTM–attention mechanism model outperformed DL for multi-classification challenges regarding terrorist activity prediction, DL outperformed the LSTM–attention mechanism model for bi-classification models.

The following are the main contributions of this paper:

• An in-depth comparative analysis of terrorist attack detection is provided in the given study;

• To accurately predict future terrorist attack activities, we propose a DL-enabled hybrid LSTM–attention mechanism model that can identify the type of attack and type of weapon;

• To predict terrorist activities, the LSTM–attention mechanism model was trained to learn spatial features from the dataset. This allowed the model to take into account a variety of factors, including the type of weapons used and whether the type of attack was successful or not.

The sections of this paper are organized as follows: the literature review is in Section 2, the methodology is in Section 3, the results and comparisons are in Section 4, and the conclusion and avenues for future work are in Section 5.

2. Literature Review

Many studies have proposed different methods to predict and detect terrorist attacks for different purposes. Regarding the literature mentioned below, we chose the highlighted studies to illustrate the prediction and prevention of terrorist attacks using machine learning and deep learning algorithms only. In a variety of disciplines, including the classification and prediction of terrorist attack features, machine learning techniques including Support Vector Machines (SVM), Artificial Neural Networks (ANN), Naïve Bayes (NB), and Random Forest have demonstrated efficacy in predictive tasks [28]. India’s terrorism problems were examined by simulating terrorist organizations’ actions using popular machine learning methods such as J48, IBK, and Naive Bayes, as well as an ensemble method that involved voting [29].

In 2017, Mo et al. [30] proposed an early warning system by using data mining techniques on a global terrorism dataset for the prediction of terrorist events and different machine learning techniques like Support Vector Machine, Naïve Bayes, and Logistic Regression. They further used two feature methods: Maximal relevance (Max-Relevance) and Minimal-redundancy maximal-relevancy (mRMR) for the improved accuracy of classification results. The proposed methodology achieved 78.41% precision in detecting early warnings of terrorist activity. The successful experiment model showed a high error rate [30].

In 2018, Alhamdani et al. [31] proposed a deep learning-based recommended system using a global terrorist dataset that helped them to detect how online media and social media are used by terrorists for the dissemination of their data. They applied multiple social alarms on different social media platforms to detect terrorist activities by considering some factors and keywords that helped them to create a decision tree against those factors to detect whether any terrorist activity was being performed or not. The experimental results showed the successful detection of terrorist activities that were spreading through social media [31].

In 2019, Kumar et al. [32] proposed a research work based on terrorist attacks from 1970 to 2015 by using a comprised dataset named terrorist attack. They applied different machine learning algorithms to discover knowledge from the dataset like random foresting, Naïve Bayes, decision trees, Lazy Classification, Multilayer Perceptron, Multiclass Classifier, etc. The main purpose of this classification was to detect whether any organization took responsibility for an attack or whether it was an anonymous attack. Their proposed system showed accuracies up to 90–95% but, due to the big size of the dataset, they chose only a few classes, which showed biased results [32].

In 2020, Semmelbeck et al. [33] proposed a model to extract features from a dataset of terrorist attacks at environmental and organizational levels using machine learning models of random forests. Then, they used inductive research design to examine patterns in the criminal behavior of terrorist groups to predict the potential risks of classification. They examined the accuracy of the model with bootstrap cross-validation and measured the factors for risks of attack. However, due to the imperfections in the data, the model showed biased results [33].

In 2020, Uddin et al. [4] proposed a model using deep learning to explore the behavior of terrorist activities. Five different deep learning models were applied to find different terrorist activities, whether an attack was successful or not, whether it was a suicide attack or not, the type of weapon, etc. The deep learning models were ensemble with deep learning models to evaluate the high performance of the models, which was recorded as 95%, but the experimental data and comparison results are not mentioned in the work [4].

In 2021, Luo et al. [34] proposed a model that helped to identify crucial indicators for the prediction of the risk of terrorist attacks and a method for the reduction of redundant information to understand the characteristics of terrorist attacks. By using the random forest method, they predicted the risk of terrorist attacks using the model [34].

In 2022, Bridgelall et al. [35] proposed a model using the natural language processing (NLP) technique to extract features from the narration of terrorists to identify terrorist attacks by using eleven different machine learning models on a dataset called terrorist attack. They applied all these machine learning models to the prediction of different narrative text fields to make the results of the model more efficient. By using different models, the proposed system acquired an accuracy of 95% but, due to the big size of the dataset, the model types were still unable to represent the whole dataset [35].

In 2022, Saidi et al. [36] proposed a model to learn temporal features from the Global Terrorist Dataset using deep learning algorithms like convolutional neural networks (CNNs) and long short-term memory (LSTM). This model aims to predict future attacks of terrorists and the behaviors of terrorists mentioned in the dataset. With the bi-classification of the proposed model, the model achieved an accuracy of 96%; however, the computation complexities were very high due to the complex structure and large size of the dataset [36].

In 2022, Arifin et al. [37] proposed a technique in which they utilized deep learning techniques to research more methods for the prediction of terrorist attacks. They used tweets to filter terrorist attack activity. For the first span, they took 100,000 tweets to run the model and extract attacking features. The model applied deep learning and two machine learning models—Naïve Bayes and K Nearest Neighbor (KNN)—to predict the results from the tweets with high accuracy, recorded as 68%; however, there is still a need to improve the system [37].

In 2023, Wang et al. [2] proposed a method of using unstructured, hidden information patterns inside a terrorist attack dataset to find the continuous and discrete features of data using machine learning algorithms. For feature selection, they used the random forest method by applying a threshold value, and then they used the XGBoost model to predict the type of terrorist attack. By using the terrorist attack dataset, they achieved a model accuracy of up to 96% but, due to the big size of the dataset, they only focused on the portions that were not folded easily [2].

In 2023, Shinde et al. [38] proposed a model that aims to provide the user with a visualization of many characteristics of terrorist attacks by using data from a dataset called global terrorism dataset. By using two different methods, they completed a graph embedding process. Then, data extracted were fed to seven different machine learning algorithms for data preprocessing and feature extraction to gain the best possible results. Hence, the classification algorithm received an accuracy of around 90%; however, the system is still under development and the authors aim to attain better results [38].

In 2023, Lamptey et al. [39] proposed a study to predict terrorist attack activity based on a deep learning model. The proposed solution showed that the new methodology was designed and implemented using SHapley Additive exPlanations (SHAP) to verify how deep learning models arrived at their decisions, which influenced the prediction results and gained an accuracy of 98% using the dataset Global Terrorist Attack [39].

In 2023, Hou et al. [1] proposed a model using a deep learning algorithm called an attention mechanism and spatial–temporal multi-graph convolutional networks to predict casual terrorist attacks at the regional level. Spatial–temporal graphs helped to generate time series information for historical data, plotting information of attack prediction using the dataset global terrorist attack and achieving high accuracy [1].

In 2022, Wang et al. [40] proposed a deep learning-based system for social computing problems using information coding technology. For this purpose, the authors designed the TOP-K algorithm to visualize the risk of terrorism by optimizing the CNN model to extract the features of the risk level. They also provided the research idea to explore counterterrorism assessment and verified the feasibility and accuracy of the model. By using the global counter-terrorism dataset, high accuracy was recorded [40].

Table 1 shows a comparative analysis of the literature in tabular form. By performing a comparative analysis of the literature, multiple successful studies were identified on terrorist attack identification and prevention, but all of these studies are lacking in results for computational complexities such as increasing data volumes or ambiguous classifiers. Moreover, there are no evident results shown by these studies that can provide details on accurate results for terrorism activities so far. Thus, in this research study, we introduce a deep learning-based methodology to predict the weapon type and attack type of an event based on learned underlying patterns within data.

3. Methodology

3.1. Dataset

The Global Terrorism Dataset (GTD) [1] is a comprehensive dataset that is extensively utilized in the identification of terrorist activities. It encompasses relevant information regarding terrorist attacks and their operations between 1970 and 2020. The National Consortium for the Study of Terrorism and Response to Terrorism (START) generated this gigantic dataset, which highlights 34 characteristics and contains over 18,100 distinct instances of terrorism. Furthermore, during the analysis of the dataset, we found the data to be highly imbalanced; therefore, we used the Synthetic Minority Over-Sampling Technique (SMOTE) to avoid biased predictions [41]. Using the SMOTE technique, we ensured that all classes had equal representation, i.e., 30,000 data samples for both cases, i.e., the prediction of attack type and weapon type. This experimental work was performed on the factors listed below:

3.1.1. Weapon Type

This factor represented the type of weapons used in terrorist attacks. The original dataset contained twelve types of weapons: Explosives, Firearms, Incendiary, Melee, Chemical, Vehicle, Sabotage Equipment, Biological, Fake Weapons, Radiological, Unknown, and Other. Some of the classes contained few samples; therefore, we merged Chemical, Vehicle, Sabotage Equipment, Biological, Fake Weapons, Radiological, Unknown, and Other into one category and named it Other. Further details about the class-wise distribution of the dataset are shown in Table 2.

3.1.2. Attack Type

This factor represented general methods and the types of attacks being conducted. The original dataset mentioned 9 different types of terrorist attacks: Assassination, Armed Assault, Bombing/Explosion, Hijacking, Hostage-taking (Barricade Incident), Hostage taking (Kidnapping), Facility/Infrastructure Attack, Unarmed Assault, and Unknown. Some classes, namely, Unknown, Unarmed Assault, Hostage-taking (Barricade Incident), and Hijacking contained fewer samples; therefore, we merged them into one class and named it Other. All the classes were balanced and contained 30,000 samples. Details about the class-wise instances are shown in Table 3.

3.2. Model Workflow

In this work, we used a deep learning model to classify terrorist attacks. Since deep learning is designed to be used with large datasets, it can be used to investigate underlying patterns and properly train a system. We present a novel deep-learning model that can be used to classify and predict different kinds of features associated with terrorist activity.

Considering its complexity, it was initially challenging to choose an efficient deep learning model for the proposed work; we conducted several tests on various deep learning models to determine which model was the most effective in terms of accurately predicting terrorist attacks. By fine-tuning various parameters of various models, we were able to compare the results and select the optimal parameters to obtain the best results.

The goal of this proposed work was to create a system that offers high accuracy for predicting terrorist attacks before they occur, enabling law enforcement authorities to take appropriate steps to prevent catastrophic actions. Thus, the purpose of this work was to investigate how terrorist actions are classified for the emergency management of terrorist attacks by taking into account the factors of weapon type and attack type. The target variables of the proposed system were run independently for each target variable (weapon type and attack type).

3.2.1. Deep Learning

Deep learning (DL) is a branch of artificial intelligence (AI) and machine learning that focuses on modeling and solving complicated problems via artificial neural networks [42]. These neural networks, which are made up of networked nodes (neurons) that analyze and alter data, are modeled after the architecture and operations of the human brain [43]. Numerous fields, such as high-frequency financial applications or weather and climate prediction, have used deep learning for time series prediction or categorization [44]. Considering this, the current work proposes a new conceptual framework based on forecasting algorithms, time series, and meta-graphs, which are frequently used in the disciplines of machine and deep learning, to bridge the gap between artificial intelligence and terrorist research. With the help of the most recent optimization techniques, the deep learning algorithm is capable of producing realistic predictions and classifications by identifying patterns in the large amount of data made available by the GTD [31].

3.2.2. Convolution Neural Networks

Convolutional neural networks (CNNs) are widely employed in computer vision applications, performing classification and prediction tasks on pictures and videos. CNNs primarily use 2-dimensional data, such as pictures and movies, to perform such tasks [45]. CNNs may, however, also process uni-dimensional information, such as time series information. Convolutional hidden layers in 1D-CNNs apply and slide a filter over a sequence, often interpreted as a non-linear alteration of the input [46].

3.2.3. Long Short-Term Memory

One of the algorithmic standards for research problems involving time series data or, more broadly, sequences, is Recurrent Neural Networks (RNNs). We used long short-term memory (LSTM) in the current study [47]. To address the widely discussed problem of vanished (or exploding) gradients in traditional RNNs, LSTM networks have been developed. They contain memory cells that enable an algorithm to learn long-term dependencies in the data more effectively by maintaining information in memory for extended periods. To prevent over-fitting, we used LSTM networks with dropout, which have a two-hidden-layer structure [48].

3.2.4. Attention Mechanism

In the field of deep learning, attention is one of the most explored ideas for challenges like picture captioning and neural machine translation. The idea of the attention mechanism as a whole can be better explained by several supporting concepts, including context vectors, hidden states, encoders, decoders, and Seq2Seq models [49]. Attention is the ability to concentrate on and pay closer attention to a particular aspect of the input problem [50]. The DL-based attention method likewise relies on focusing your attention and giving particular aspects of a problem more attention while processing pertinent facts [51].

We propose a hybrid CNN-LSTM model in this research for predicting terrorist activities. The LSTM layer extracts context-dependent features to improve the accuracy of the CNN model, which learns the local features of a dataset. The proposed architecture of the system is represented in Figure 1; it consists of an input-filtered dataset, a fully connected layer, an LSTM layer, a 1D convolution layer, and a soft-max function for the classification task. The working of the proposed system consists of two deep neural network blocks, each of which has five hidden layers and a ReLU activation function to achieve all of the previously stated goals. In the case of binary and multiclass evaluations, Sigmoid and SoftMax are also utilized.

3.3. Proposed Architecture

Initially, in loading and preprocessing the data, we divided the features and the target labels, denoted by X and Y, respectively. The process of feature selection was used to minimize the number of features. In this instance, we utilized SelectKBest to determine the top ‘10’ features according to the F-statistic of the ANOVA.

Then, in the second step, the splitting and normalization of the dataset were completed by using the train test split function; the dataset was split into training, validation, and test sets for normalization. In contrast, StandardScaler was used to scale the features to have a zero mean and unit variance as part of the data standardization process.

After the splitting and normalization of the dataset, training, validation, and testing of the datasets were conducted to reshape data for the CNN using LSTM so that they fit the model’s input shape. They were reshaped in this particular case to function with 1D convolutional neural network (CNN) architecture, providing dimensions (samples, features, and 1), where the network layers for categorical categorization comprised several CNNs, an LSTM layer, and a SoftMax. The LSTM layer, attention layers, max-pooling layers, and convolutional layers contributed to the architecture of the model when it was built. The convolutional layers were intended to capture different amounts of information and employ various filter sizes. The output of the convolutional layers was subsequently processed by the LSTM layer. To utilize the most optimal features, the attention layer was used. The model was compiled with the specified loss function (cross-entropy) and optimizer (Adam). We used metrics for the evaluation of accuracy. The training dataset was used to train the model and the validation dataset was used to validate the dataset. With a given batch size of 32 and epoch count of 50 for type of weapon and type of attack, the training was carried out. The validation dataset was used to evaluate the model’s performance once it had been trained.

4. Results

In this study of the accurate detection of terrorist activities, we mainly targeted the type of attack and weapon type; we employed the benchmark GTD to compare the updated DNN composed of the CNN-LSTM model with the latest studies, including the benchmark machine learning model and some deep learning models. Results were generated following a training and testing split ratio of 80:20.

We used a Python framework for the development of the proposed system using libraries Pandas, Scikit-learn, and TensorFlow. The system was trained and tested on Windows 10 with a core i7 processor and 16 GB of RAM.

4.1. Results and Performance Measures

To measure the accuracy of the results for the proposed model, we used evaluation measures using true positive, true negative, false positive, and false negative values. We measured the accuracy as an evaluation measure to obtain the results as compared to traditional models. The reasons behind using traditional evaluation measures are defined below. Briefly, TP is a true positive, TN is a true negative, FP is a false positive, and FN is a false negative. To assess accuracy, we evaluated the ratio between the prediction of the outcome and the sum of the prediction of the model’s true value, calculated using Equation (1), given below.

(1)$\mathrm{Accuracy}={\displaystyle \frac{TP+TN}{FP+FN+TP+TN}}$

Table 4 and Table 5 represent the training and testing accuracy and loss against the proposed model, also shown in the figures shown below; the green curve represents the training results, whereas the blue curve represents the testing results. Figure 2 presents the confusion matrix for attack type prediction, while Figure 3 shows the confusion matrix for weapon type prediction.

Figure 4 and Figure 5 represent the accuracy and loss of attack type prediction. Figure 4 shows the accuracy graph for attack type. The training accuracy and test accuracy were relatively close to each other during the increase in epochs during the model configuration. The accuracy of the model increased during the testing and training of the model; hence, the graph shows an upward trend. Figure 5 shows the loss graph for attack type. The training loss and test loss were relatively close to each other during the increase in epochs during the model configuration; the value of loss showed a downward trend during the testing and training of the model. In the case of both accuracy and loss, the graph generated a constant when the epoch counts reached 50 with a batch size of 32. The precision, recall, and F1-score for both class labels are reported in Table 4, and the highest testing accuracy was 97.6%, whereas the loss was 0.024. Figure 6 and Figure 7 represent the accuracy and loss of weapon type prediction, respectively.

The detailed epoch-wise performance of the model for accuracy and loss for the type of weapon used can be seen in Figure 6 and Figure 7, respectively. The proposed model was trained on various numbers of hyperparameters and under different settings.

Table 5 represents the optimal parameters for the proposed model. The model was trained over 50 epochs for type of attack and type of weapon, with a batch size of 32 using the loss function cross-entropy and Adam optimizer, Sigmoid as an activation function, and dropout of 0.5; the obtained results can be seen in Table 1. A maximum testing accuracy of 98.12% was obtained for attack type prediction, while 97.6% was obtained for weapon type prediction.

4.2. Comparison with Existing Approaches

In this section, we present a detailed comparative analysis of different approaches used on the GTD by researchers. Table 6 shows a comparison of the proposed model with existing approaches.

A comparison of studies is shown in Table 6 using the same dataset for different factors and recording significant results. In [52], Pan at al. 2021 proposed a model for a classification framework based on ensemble learning for classifying and predicting using the GTD and achieved 97.1% accuracy; however, further improvement in the performance and accuracy of machine learning algorithms is required.

In [36], the authors proposed a hybrid CNN-LSTM and DL framework for predicting future terrorist activities using GTD and achieved 99.2% accuracy; however, the model did not incorporate the local features of the dataset. In [4], Uddin et al. proposed a model for the prediction of future terrorist attacks by using five deep learning models using GTD and achieved 95% accuracy; however, not all the models were able to make predictions of future attacks with high accuracy. In [53], Jiang et al. proposed a deep-learning and multi-level framework for understanding the behavior of terrorist groups using the GTD, but the results and implementation details were not mentioned. In [54], Campedelli et al. proposed temporal meta-graphs and a deep learning model for the forecasting of future terrorist targets using the GTD and achieved 95% accuracy; however, most of the regional attacks predicted were based on predictions. In [55], Campedelli et al. a multi-modal networks is proposed to reveal patterns of operational similarity among terrorist organizations using the GTD, achieving an accuracy of 96.5%; however, the experimental results were for specific years (2012–2018). In [50], Nayak et al. proposed a deep-learning model using an attention-based hybrid CNN and Bi-LSTM model to learn future terrorist targets using the GTD, achieved 97.2% accuracy; however, in the proposed model, not all the models showed learning rates accurately. In [37], Arifin et al. proposed a deep learning model to predict terms used by terrorists to pre-plan attacks using real-time Twitter tweets from Rapid Miner using the GTD and achieved 89% accuracy; however, the details of the experiment were not shown. In [56], Bangerter et al. proposed a network strategy that was developed to measure the terrorist network in a region using the GTD [59] and achieved good accuracy in the results; however, the network became complex as the depth of the nodes increased. Hence, we propose a hybrid CNN-LSTM model for the detection of terrorist attacks using the GTD and achieved 98.12% accuracy with this model; however, more target factors can be chosen from the dataset in the future.

Hence, compared to all the benchmark methodologies, which used various models for the detection of terrorist attacks, our proposed model, the LSTM–attention mechanism model was trained to learn the spatial features of the GTD. This allowed the model to take into account a variety of factors, including the type of weapons used and whether the type of attack was successful or not, and showed the best performance in terms of both training and testing.

5. Conclusions

Terrorism is one of the most important critical threats to mankind so far and is affecting the quality of life at numerous levels. To forecast future terrorist activities for the prevention of multiple aspects like human life loss, political crises, law and governance crises, and many others, we proposed a hybrid DL model in this study. Our proposed hybrid DL model was composed of a CNN, LSTM, and attention mechanism to record the temporal features of data, using the benchmark dataset the GTD. From the given dataset, we predicted the following: type of attack and type of weapon. We compared our proposed model with nine benchmark studies using AI and ML models on the same dataset and found that our proposed hybrid model achieved the highest accuracy results, with 98.12% and 97.6% for attack type and weapon type prediction, respectively.

In the future, the proposed hybrid approach will perform effectively with the same datasets. However, it currently performs poorly on more varied datasets. To solve this problem, we are keen to look into various transfer learning-based networks and efficient nets or ResNets to enhance the network’s performance more significantly.

Author Contributions

Conceptualization, E.U.H.Q. and M.H.F.; data curation, E.U.H.Q., M.H.F., T.Z., M.I. and I.A.; formal analysis, E.U.H.Q. and M.H.F.; funding acquisition, E.U.H.Q.; methodology, E.U.H.Q. and M.H.F.; project administration, E.U.H.Q.; resources, E.U.H.Q.; software, E.U.H.Q. and M.H.F.; supervision, E.U.H.Q.; validation, E.U.H.Q., M.H.F., T.Z., M.I. and I.A.; visualization, E.U.H.Q., M.H.F., T.Z., M.I. and I.A.; writing—original draft, E.U.H.Q. and M.H.F.; writing—review and editing. E.U.H.Q., M.H.F., T.Z., M.I. and I.A. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

The data that support the findings of this study are available in the Global Terrorism Database (GTD) repository, https://www.start.umd.edu/gtd/ (accessed on 2 October 2024).

Acknowledgments

The authors would like to express their deep thanks to the Vice Presidency for Scientific Research at Naif Arab University for Security Sciences for their kind encouragement of this work.

Conflicts of Interest

The authors declare no conflicts of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. Proposed architecture diagram.

Figure 2. Attack type confusion matrix.

Figure 3. Weapon type confusion matrix.

Figure 4. Attack type accuracy graph.

Figure 5. Attack type loss graph.

Figure 6. Weapon type accuracy graph.

Figure 7. Weapon type loss graph.

Table 1

Comparative analysis of literature review.

Ref.	Proposed Method	Dataset	Results	Limitations
[1]	They proposed a deep learning method to predict time series information of terrorist attacks.	Global Terrorist Attack	-	Experimental results are not compared with other methodologies.
[2]	They identified unstructured data from a dataset to identify terrorist attack risk.	Terrorist Attack	Accuracy 96.34%	There are no details of the experiment mentioned.
[4]	They detected different behaviors of terrorist attacks using different deep learning techniques.	Global Terrorist Attack	Accuracy 95%	The system performs complex computations due to the involvement of different deep-learning models.
[30]	They proposed a model for the early detection of terrorist attacks using an early warning detector.	Global Terrorist Attack	Precision 78.41%	The proposed system has a higher error rate in data classification.
[31]	They detected how online media and social media are used by terrorists for the dissemination of their data.	Global Terrorist Attack	-	There are no details of the experiment mentioned.Experimental results are not mentioned and compared with other methodologies.
[32]	They detected whether any organization took responsibility for an attack or whether it was an anonymous attack.	Terrorist Attack	Accuracy 90–95%	Because of the big size of the dataset, the experimental work was performed on a small portion of the data, so the results were biased.
[33]	They identified the risk of terrorist attacks in an environment and organization using machine learning algorithms.	-	-	The results were biased due to imperfections in the dataset.
[34]	They identified crucial indicators for the prediction of the risk of terrorist attacks.	Self-Generated	-	There are no details of the experiment mentioned. Experimental results are not mentioned and compared with other methodologies.
[35]	They detected the narration of terrorists for the prevention of terrorist attacks using NLP and eleven machine learning models.	Terrorist Attack	Accuracy 95%	Due to the large size of the dataset, the data were not fully used and the results were not well classified.
[36]	They detected terrorist attack behaviors for future attacks using a deep learning model from GAD.	Global Terrorist Attack	Accuracy 96%	The system has computational complexity due to the large size of the dataset and the complex structure of the deep learning model.
[37]	They proposed a system to predict the comprehensive behavior of terrorist attacks.	Terrorist Attack	Accuracy 68%	More improvements to the proposed system are needed for better results.
[38]	They proposed a model in which users can visualize the risk of terrorist attacks using machine learning models.	Global Terrorist Attack	Accuracy 90%	The system is still under development and the authors aim to attain better results.
[39]	They predicted terrorist attacks using deep learning algorithms and the SHAP method.	Global Terrorist Attack	Accuracy 96%	The detailed workings of the model are not mentioned.
[40]	They proposed a model to detect early warnings for counterterrorism with the help of a deep learning model.	Global Counter-Terrorism	-	Experimental results are not compared with other methodologies.

Table 2

Global Terrorism Dataset (GTD) weapon types.

Weapon Type	Instances
Explosives	30,000 per class
Firearms
Incendiary
Melee
Other

Table 3

Global Terrorism Dataset (GTD) attack types.

Attack Type	Instances
Bombing/Explosion	30,000 per class
Armed Assault
Assassination
Hostage-taking (Kidnapping)
Facility/Infrastructure Attack
Other

Table 4

Model results.

Class Label	Accuracy		Loss		Precision	Recall	F1-Score
	Training (%)	Testing (%)	Training	Testing
Attack Type	99.2	98.12	0.008	0.188	98.12	98.12	98.11
Weapon Type	98.1	97.6	0.019	0.024	97.60	97.60	97.59

Table 5

Model parameters.

Settings	Parameters
Epochs:	Attack Type 50Weapon Type 50
Loss function:	Cross-entropy
Optimizer:	Adam
Activation function:	Sigmoid
Batch size:	32
Dropout:	0.5

Table 6

Comparison of the proposed model with existing approaches.

Ref.	Methodology	Dataset	Limitations	Accuracy
[4](Uddin, et al.), 2020	Future terrorist attack prediction using five deep learning models.	Global Terrorism Database (GTD)	The models are not able to make predictions of future attacks with high accuracy.	95%
[36](Saidi, et al.), 2022	Hybrid CNN-LSTM and DL framework for predicting future terrorist activities.	Global Terrorism Database (GTD)	The model does not incorporate the local features of the dataset.	99.2%
[37]	Deep learning model to predict terms used by terrorists to pre-plan an attack on real-time Twitter tweets from Rapid Miner.	Global Terrorism Database (GTD)	The results and implementation details are not mentioned.	89%
[51]	A deep learning model using an attention-based hybrid CNN and BiLSTM model to learn future terrorist targets.	Global Terrorism Database (GTD)	In the proposed model, not all the models show learning rates accurately.	97.2%
[52]	Classification framework based on ensemble learning for classifying and predicting.	Global Terrorism Database (GTD)	Further improvements in the performance and accuracy of machine learning algorithms are required.	97.1%
[53]	A deep-learning and multi-level framework for understanding the behavior of terrorist groups.	Global Terrorism Database (GTD)	The results and implementation details are not mentioned.	-
[54]	Temporal meta-graphs and a deep learning model for the forecasting of future terrorist targets.	Global Terrorism Database (GTD)	Most regional attacks are predicted based on predictions.	95%
[55]	A multi-modal network reveals patterns of operational similarity among terrorist organizations.	Global Terrorism Database (GTD)	The experimental results were for specific years: 2012–2018.	96.5%
[56]	Network strategy developed to measure the terrorist network in a region.	Global Terrorism Database (GTD)	The network becomes complex as the depth of the nodes increases.	-
[57]	Hybrid LSTM–CNN	Legitimate and phishing websites	LSTM–CNN and LSTM models performed slightly lower in accuracy	97.6%
[58]	Hybrid Deep Learning-Based Network Intrusion Detection System	CICIDS-2018	The deep-layered architecture combining CNN and RNN is resource-intensive	98.9%
Proposed Study	Hybrid CNN-LSTM model for the detection of terrorist attacks.	Global Terrorism Database (GTD)	More target factors can be chosen from the dataset.	Attack Type: 98.12%Weapon Type: 97.6%

References

1. Hou, Z.; Zhou, Y.; Wu, X.; Bu, F. Attention-Based Spatial–Temporal Multi-Graph Convolutional Networks for Casualty Prediction of Terrorist Attacks. Complex Intell. Syst.; 2023; 9, pp. 6307-6328. [DOI: https://dx.doi.org/10.1007/s40747-023-01037-z]

2. Wang, Y. Machine Learning Based Terrorist Attacks Prediction Algorithm. Appl. Comput. Eng.; 2023; 2, pp. 391-397. [DOI: https://dx.doi.org/10.54254/2755-2721/2/20220537]

3. Miller, R. Book Review: Robert J. Bunker (Ed.), Networks, Terrorism and Global Insurgency, London and New York, Routledge, 2005; XXVI + 211 pp.; £26.99 pbk; ISBN 0415385946 Brynjar Lia, Globalisation and the Future of Terrorism: Patterns and Predictions, London and New York, Routledge, 2005; XVII + 259 pp.; £21.99 pbk; ISBN 0415402965. J. Contemp. Hist.; 2009; 44, pp. 373-376. [DOI: https://dx.doi.org/10.1177/00220094090440020820]

4. Uddin, M.I.; Zada, N.; Aziz, F.; Saeed, Y.; Zeb, A.; Ali Shah, S.A.; Al-Khasawneh, M.A.; Mahmoud, M. Prediction of Future Terrorist Activities Using Deep Neural Networks. Complexity; 2020; 2020, 1373087. [DOI: https://dx.doi.org/10.1155/2020/1373087]

5. Futia, G.; Vetrò, A. On the Integration of Knowledge Graphs into Deep Learning Models for a More Comprehensible AI-Three Challenges for Future Research. Information; 2020; 11, 122. [DOI: https://dx.doi.org/10.3390/info11020122]

6. Wen, D.; He, M.; Wang, Y.; Zhang, Y. Forecasting Stock Market Realized Volatility: The Role of Global Terrorist Attacks. Appl. Econ.; 2023; 55, pp. 2551-2566. [DOI: https://dx.doi.org/10.1080/00036846.2022.2103503]

7. Li, Y.; Wu, Y. Research on the Impact of Crisis Events on Urban Development—A Case Study of Kunming Railway Station Terrorist Attack. Proceedings of the 20th International Symposium on Advancement of Construction Management and Real Estate; Springer: Singapore, 2017.

8. Vasilopoulos, P. Terrorist Events, Emotional Reactions, and Political Participation: The 2015 Paris Attacks. West Eur. Politics; 2018; 41, pp. 102-127. [DOI: https://dx.doi.org/10.1080/01402382.2017.1346901]

9. Hu, X.; Lai, F.; Chen, G.; Zou, R.; Feng, Q. Quantitative Research on Global Terrorist Attacks and Terrorist Attack Classification. Sustainability; 2019; 11, 1487. [DOI: https://dx.doi.org/10.3390/su11051487]

10. Chauvelin, L.; Gindt, M.; Olliac, B.; Robert, P.; Thümmler, S.; Askenazy, F. Emergency Organization of Child Psychiatric Care Following the Terrorist Attack on July 14, 2016, in Nice, France. Disaster Med. Public Health Prep.; 2019; 13, pp. 144-146. [DOI: https://dx.doi.org/10.1017/dmp.2018.51]

11. Weeratna, J.; Tennakoon, A. Management of The Dead and Missing People in the Easter Bombings in Colombo, Sri Lanka. J. Clin. Health Sci.; 2021; 6, 59. [DOI: https://dx.doi.org/10.24191/jchs.v6i1(Special).13163]

12. Tallach, R.; Einav, S.; Brohi, K.; Abayajeewa, K.; Abback, P.S.; Aylwin, C.; Batrick, N.; Boutonnet, M.; Cheatham, M.; Cook, F. et al. Learning from Terrorist Mass Casualty Incidents: A Global Survey. Br. J. Anaesth.; 2022; 128, pp. e168-e179. [DOI: https://dx.doi.org/10.1016/j.bja.2021.10.003] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34749991]

13. Feng, Y.; Wang, D.; Yin, Y.; Li, Z.; Hu, Z. An XGBoost-Based Casualty Prediction Method for Terrorist Attacks. Complex Intell. Syst.; 2020; 6, pp. 721-740. [DOI: https://dx.doi.org/10.1007/s40747-020-00173-0]

14. Song, Y.; Chen, B.; Wang, X.Y. Cryptocurrency Technology Revolution: Are Bitcoin Prices and Terrorist Attacks Related?. Financ. Innov.; 2023; 9, 29. [DOI: https://dx.doi.org/10.1186/s40854-022-00445-3] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36712148]

15. Zhu, R.; Hu, X.; Li, X.; Ye, H.; Jia, N. Modeling and Risk Analysis of Chemical Terrorist Attacks: A Bayesian Network Method. Int. J. Environ. Res. Public Health; 2020; 17, 2051. [DOI: https://dx.doi.org/10.3390/ijerph17062051]

16. LaFree, G.; Dugan, L. Introducing the Global Terrorism Database. Terror. Political Violence; 2007; 19, pp. 181-204. [DOI: https://dx.doi.org/10.1080/09546550701246817]

17. Petroff, V.B.; Bond, J.H.; Bond, D.H. Using Hidden Markov Models to Predict Terror before It Hits (Again). Handbook of Computational Approaches to Counterterrorism; Springer: New York, NY, USA, 2013; ISBN 9781461453116

18. Meng, X.; Nie, L.; Song, J. Big Data-Based Prediction of Terrorist Attacks. Comput. Electr. Eng.; 2019; 77, pp. 120-127. [DOI: https://dx.doi.org/10.1016/j.compeleceng.2019.05.013]

19. Rai, A.K.; Kumar, S. Identifying the Leaders and Main Conspirators of the Attacks in Terrorist Networks. ETRI J.; 2022; 44, pp. 977-990. [DOI: https://dx.doi.org/10.4218/etrij.2021-0239]

20. Lu, P.; Li, M.; Zhang, Z. The Crowd Dynamics under Terrorist Attacks Revealed by Simulations of Three-Dimensional Agents. Artif. Intell. Rev.; 2023; 56, pp. 13103-13125. [DOI: https://dx.doi.org/10.1007/s10462-023-10452-0]

21. Chatterjee, S.; Abkowitz, M.D. A Methodology for Modeling Regional Terrorism Risk. Risk Anal.; 2011; 31, pp. 1133-1140. [DOI: https://dx.doi.org/10.1111/j.1539-6924.2010.01565.x]

22. Al-Dahash, H.; Kulatunga, U.; Allali, B. Factors Affecting Risk Perception during Terrorist Attacks. Int. J. Disaster Risk Reduct.; 2022; 73, 102870. [DOI: https://dx.doi.org/10.1016/j.ijdrr.2022.102870]

23. Kebir, O.; Nouaouri, I.; Rejeb, L.; Ben Said, L. ATiPreTA: An Analytical Model for Time-Dependent Prediction of Terrorist Attacks. Int. J. Appl. Math. Comput. Sci.; 2022; 32, pp. 495-510. [DOI: https://dx.doi.org/10.34768/amcs-2022-0036]

24. Baby, A.; Sruthi, A. Machine learning models for prediction of terrorist attacks: A comparative analysis in terrorism prone regions. Proceedings of the 2023 Annual International Conference on Emerging Research Areas: International Conference on Intelligent Systems (AICERA/ICIS); Kanjirapally, India, 16–18 November 2023; pp. 1-6.

25. de Oliveira, R.A.; Bollen, M.H.J. Deep Learning for Power Quality. Electr. Power Syst. Res.; 2023; 214, 108887. [DOI: https://dx.doi.org/10.1016/j.epsr.2022.108887]

26. Huamaní, E.L.; Alicia, A.M.; Roman-Gonzalez, A. Machine Learning Techniques to Visualize and Predict Terrorist Attacks Worldwide Using the Global Terrorism Database. Int. J. Adv. Comput. Sci. Appl.; 2020; 11, pp. 562-570. [DOI: https://dx.doi.org/10.14569/IJACSA.2020.0110474]

27. Chen, J.H.; Asch, S.M. Machine Learning and Prediction in Medicine—Beyond the Peak of Inflated Expectations. N. Engl. J. Med.; 2017; 376, pp. 2507-2509. [DOI: https://dx.doi.org/10.1056/NEJMp1702071]

28. Verma, C.; Verma, V. Predictive Modeling of Terrorist Attacks Using Machine Learning. Int. J. Pure Appl. Math.; 2018; 119, pp. 49-61.

29. Gundabathula, V.T.; Vaidhehi, V. An Efficient Modelling of Terrorist Groups in India Using Machine Learning Algorithms. Indian J. Sci. Technol.; 2018; 11, pp. 1-10. [DOI: https://dx.doi.org/10.17485/ijst/2018/v11i15/121766]

30. Mo, H.; Meng, X.; Li, J.; Zhao, S. Terrorist Event Prediction Based on Revealing Data. Proceedings of the 2017 IEEE 2nd International Conference on Big Data Analysis, ICBDA 2017; Beijing, China, 10–12 March 2017.

31. Alhamdani, R.S.; Abdullah, M.N.; Sattar, I.A. Recommender System for Global Terrorist Database Based on Deep Learning. Int. J. Mach. Learn. Comput.; 2018; 8, 6. [DOI: https://dx.doi.org/10.18178/ijmlc.2018.8.6.747]

32. Kumar, V.; Mazzara, M.; Messina, A.; Lee, J.Y. A Conjoint Application of Data Mining Techniques for Analysis of Global Terrorist Attacks: Prevention and Prediction for Combating Terrorism. Proceedings of 6th International Conference in Software Engineering for Defence Applications; Springer: Cham, Switzerland, 2020.

33. Semmelbeck, J.; Besaw, C. Exploring the Determinants of Crime-Terror Cooperation Using Machine Learning. J. Quant. Criminol.; 2020; 36, pp. 527-558. [DOI: https://dx.doi.org/10.1007/s10940-019-09421-0]

34. Luo, L.; Qi, C. An Analysis of the Crucial Indicators Impacting the Risk of Terrorist Attacks: A Predictive Perspective. Saf. Sci.; 2021; 144, 105442. [DOI: https://dx.doi.org/10.1016/j.ssci.2021.105442]

35. Bridgelall, R. An Application of Natural Language Processing to Classify What Terrorists Say They Want. Soc. Sci.; 2022; 11, 23. [DOI: https://dx.doi.org/10.3390/socsci11010023]

36. Saidi, F.; Trabelsi, Z. Corrigendum to “A hybrid deep learning-based framework for future terrorist activities modeling and prediction” [Egypt. Inform. J. 23 (3) 2022, 437–446]. Egypt. Inform. J.; 2023; 24, pp. 215-216. [DOI: https://dx.doi.org/10.1016/j.eij.2023.03.002]

37. Arifin, V.; Jallow, F.B.; Lubis, A.; Bahaweres, R.B.; Rofiq, A.A. Using Deep Learning Model to Predict Terms Use by Terrorist to Pre-Plan an Attack on a Real-Time Twitter Tweets from Rapid Miner. Proceedings of the 2022 10th International Conference on Cyber and IT Service Management, CITSM 2022; Yogyakarta, Indonesia, 20–21 September 2022.

38. Shinde, S.; Khoje, S.; Raj, A.; Wadhwa, L.; Shaikh, A.S. Artificial Intelligence Approach for Terror Attacks Prediction through Machine Learning. Multidiscip. Sci. J.; 2024; 6, 2024011. [DOI: https://dx.doi.org/10.31893/multiscience.2024011]

39. Lamptey, O.; Gegov, A.; Ouelhadj, D.; Hopgood, A.; Da Deppo, S. Neural Network Based Identification of Terrorist Groups Using Explainable Artificial Intelligence. Proceedings of the 2023 IEEE Conference on Artificial Intelligence (CAI); Santa Clara, CA, USA, 5–6 June 2023; pp. 191-192. [DOI: https://dx.doi.org/10.1109/CAI54212.2023.00090]

40. Wang, Y.; Cui, X.; He, P. Winning the War on Terror: Using “Top-K” Algorithm and CNN to Assess the Risk of Terrorists. Int. J. Inf. Technol. Web Eng.; 2022; 17, pp. 1-15. [DOI: https://dx.doi.org/10.4018/IJITWE.288038]

41. Chawla, N.V.; Bowyer, K.W.; Hall, L.O.; Kegelmeyer, W.P. SMOTE: Synthetic minority over-sampling technique. J. Artif. Int. Res.; 2002; 16, pp. 321-357. [DOI: https://dx.doi.org/10.1613/jair.953]

42. Ali, S.; Ali, H.H.; Qayyum, S.; Sohail, F.; Tahir, F.; Maqsood, S.; Adil, M. Multi-Agent System Using Scrum Methodology for Software Process Management. Intelligent Technologies and Applications; Springer: Singapore, 2019.

43. Janiesch, C.; Zschech, P.; Heinrich, K. Machine Learning and Deep Learning. Electron. Mark.; 2021; 31, pp. 685-695. [DOI: https://dx.doi.org/10.1007/s12525-021-00475-2]

44. Bui, D.T.; Tsangaratos, P.; Nguyen, V.T.; Van Liem, N.; Trinh, P.T. Comparing the Prediction Performance of a Deep Learning Neural Network Model with Conventional Machine Learning Models in Landslide Susceptibility Assessment. Catena; 2020; 188, 104426. [DOI: https://dx.doi.org/10.1016/j.catena.2019.104426]

45. Mayerich, D.; Sun, R.; Guo, J. Deep Learning. Microscope Image Processing; 2nd ed. Academic Press: Cambridge, MA, USA, 2022; ISBN 9780128210499

46. Ismail Fawaz, H.; Forestier, G.; Weber, J.; Idoumghar, L.; Muller, P.A. Deep Learning for Time Series Classification: A Review. Data Min. Knowl. Discov.; 2019; 33, pp. 917-963. [DOI: https://dx.doi.org/10.1007/s10618-019-00619-1]

47. Hochreiter, S.; Schmidhuber, J. Long Short-Term Memory. Neural Comput.; 1997; 9, pp. 1735-1780. [DOI: https://dx.doi.org/10.1162/neco.1997.9.8.1735]

48. Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; Salakhutdinov, R. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. J. Mach. Learn. Res.; 2014; 15, pp. 1929-1958.

49. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, Ł.; Polosukhin, I. Attention Is All You Need. Proceedings of the 31st Conference on Neural Information Processing Systems (NIPS 2017); Long Beach, CA, USA, 4–9 December 2017.

50. Soydaner, D. Attention Mechanism in Neural Networks: Where It Comes and Where It Goes. Neural Comput. Appl.; 2022; 34, pp. 13371-13385. [DOI: https://dx.doi.org/10.1007/s00521-022-07366-3]

51. Nayak, N.; Rayachoti, M.; Gupta, A.M.; Prerna, G.P.; Sreenath, V.M.; Annapurna, D. Learning Future Terrorist Targets Using Attention Based Hybrid CNN and BiLSTM Model. Proceedings of the 2023 IEEE International Conference on Integrated Circuits and Communication Systems, ICICACS 2023; Raichur, India, 24–25 February 2023.

52. Pan, X. Quantitative Analysis and Prediction of Global Terrorist Attacks Based on Machine Learning. Sci. Program.; 2021; 2021, 7890923. [DOI: https://dx.doi.org/10.1155/2021/7890923]

53. Jiang, D.; Wu, J.; Ding, F.; Ide, T.; Scheffran, J.; Helman, D.; Zhang, S.; Qian, Y.; Fu, J.; Chen, S. et al. An Integrated Deep-Learning and Multi-Level Framework for Understanding the Behavior of Terrorist Groups. Heliyon; 2023; 9, e18895. [DOI: https://dx.doi.org/10.1016/j.heliyon.2023.e18895] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37636372]

54. Campedelli, G.M.; Bartulovic, M.; Carley, K.M. Learning Future Terrorist Targets through Temporal Meta-Graphs. Sci. Rep.; 2021; 11, 8533. [DOI: https://dx.doi.org/10.1038/s41598-021-87709-7]

55. Campedelli, G.M.; Cruickshank, I.; Carley, K.M. Multi-Modal Networks Reveal Patterns of Operational Similarity of Terrorist Organizations. Terror. Political Violence; 2023; 35, pp. 1065-1084. [DOI: https://dx.doi.org/10.1080/09546553.2021.2003785]

56. Bangerter, M.L.; Fenza, G.; Gallo, M.; Loia, V.; Petrone, A.; Volpe, A. Terrorist Organization Identification Using Link Prediction over Heterogeneous GNN. Hum.-Centric Comput. Inf. Sci.; 2022; 12, pp. 1-13. [DOI: https://dx.doi.org/10.22967/HCIS.2022.12.012]

57. Alshingiti, Z.; Alaqel, R.; Al-Muhtadi, J.; Haq, Q.E.U.; Saleem, K.; Faheem, M.H. A Deep Learning-Based Phishing Detection System Using CNN, LSTM, and LSTM-CNN. Electronics; 2023; 12, 232. [DOI: https://dx.doi.org/10.3390/electronics12010232]

58. Qazi, E.U.H.; Faheem, M.H.; Zia, T. HDLNIDS: Hybrid Deep-Learning-Based Network Intrusion Detection System. Appl. Sci.; 2023; 13, 4921. [DOI: https://dx.doi.org/10.3390/app13084921]

59. START (National Consortium for the Study of Terrorism and Responses to Terrorism). Global Terrorism Database 1970–2020 [Data File]. 2022; Available online: https://www.start.umd.edu/gtd (accessed on 4 April 2024).