Enhancing Financial Fraud Detection Through Chimp-Optimized Long Short-Term Memory Networks

Table 1. Comparative analysis of the existing works

Many supervised and unsupervised learning methods are employed for transaction-based fraud detection. Some of the most important are as follows:

Sudarshana et al. [11] offer an autoencoder-based classification technique for extracting variables from European credit card data. The reliability of several machine learning approaches for categorization was also evaluated using encoded features. The use of a neural network autoencoder is presented as a method for detecting fraudulent purchases made with a credit card [12]. Generally, spotting credit card fraud is seen as a problem of categorization. Esenogho et al. [13] suggests a reliable credit card fraud detection method by combining a hybrid data resampling approach with a neural network ensemble classifier.

One of the largest online marketplaces is targeted by Zhang et al. [14], which offers a fraud detection system (eFraudCom) based on competitive graph neural networks (CGNN). Trials conducted on two Taobao datasets and two public datasets demonstrated that the suggested architecture, CGNN, significantly outperformed prior baselines in detecting fraudulent activities. Incorporating sequences of transactions into a credit card fraudulent recognition system requires the use of a sequence learner, such as a Long Short-Term Memory (LSTM) network, as described in Benchaji et al. [15]. With the intention of increasing the accuracy of fraud detection on new incoming transactions, the suggested approach tries to record the historical purchasing behavior of credit card holders. To detect fraudulent activity, Daliri [16] uses an artificial neural network strategy based on the Harmony Search Algorithm. The proposed approach mines the data of both genuine and fraudulent customers for underlying trends. Since fraudulent behavior may be detected and prevented before it ever happens, the results show that the proposed system has sufficient competence in fraud detection. Various Comparative analysis of the existing works were mentioned in Table 1.

A unique fraud detection system named FraudMemory is proposed by Yang and Xu [17]. To better represent users and logs of various types in financial systems, it employs cutting-edge feature representation techniques.

In summary, the literature survey highlights diverse approaches to credit card fraud detection, with varying success. However, there are gaps in method transparency, adaptability to online fraud, and consistent reporting. This research aims to fill these gaps by proposing a transparent and adaptable fraud detection system, addressing limitations identified in existing studies. As a result, it is suggested that optimization-based FS be used to identify fraudulent online purchases.

Three different dataset types—mortgage, credit card, and insurance—are first imported. Then, using a conventional scalar-based preprocessing method, the risk variables are scaled and values assigned to help detect the difference and decrease noise. Some fuzzy rules are applied after preprocessing to retrieve the features. Following feature extraction, an optimization-based FS is carried out to choose the features of users who have produced suspicious, fraudulent, and fraudulent results. Finally, classification based on long- and short-term memory is completed. Figure 1 illustrates the Architecture of Developed Model observed in fraud detection. This highlights the standard scalar method, fuzzy based feature extraction, LSTM, chimp optimization.

1.jpg

Figure 1. Architecture of developed model

4.1 Preprocessing of data using standard scalar method

The dataset should be chosen as the initial step from the online machine learning repositories. After that, we repaired and standardized the data sets that were gathered. These datasets have incorrect values and weren't collected in a controlled setting. As a result, data preparation is an essential step in data analysis and machine learning. Data normalization is the process of standardizing a dataset in which individual risk variables have different values. Data standardization involves assigning values to risk variables that reflect the distribution of standard deviations from the mean. It adjusts the risk factor value such that classifiers with a mean of 0 and a standard deviation of 1 perform better. Standardization may be expressed mathematically as:

standardization of $X=\frac{X-\text { mean of } X}{\text { standard deviation of } X}$          (2)

The idea behind this technique is to choose a set of samples that are close together in the feature space, draw a line between the data points, and then create a new instance of the minority class somewhere along the line after the criteria have been established.

4.2 Fuzzy based future extraction

Following preprocessing, fuzzy concepts are used to extract the features. This section performs the fundamental checks for each inbound transaction on a given card. The PIN, the transaction total, and the card's expiration date are all factors.

Rule 1: Matching consistency features. To protect the security of their common phone number and avoid getting caught, fraudsters use fresh electronic payment phone numbers for a lot of their illegal activities. Consequently, we establish the rule for matching consistency features. To check whether a common phone and a pay-bound phone in a transaction record are the same, we build the matching function Match(a,b). If the new feature phone numbers are the same, they get a 0, and if they're different, a 1.

$\operatorname{match}\left(\right.$ num $_{\text {com }}$, num $\left._{\text {e-pay }}\right\}=\left\{\right.$ true num $_{\text {com }}=$ num $_{\text {e-pay }}$ false otherwise         (3)

Rule 2: The validation rule is unreliable because some suspicious transactions will be examined by the staff over the phone with the cardholder, other people will be verified as fraudulent, and still others will not be demonstrated because no one answers when the staff calls. The opinion of returning bank employees (pre_(trade_result )) as a means of checking the legitimacy of an unusual financial transaction that the system has intercepted V_f={V₀, V₁, null} where V₀ is the verification result as true transaction, V₁ is the verification result as fraud transaction and NULL is the suspicious transaction. We combine pre $_{\text {trade }_{\text {result }}}$ with is_common_ip for further verification. with $i s_{\text {common }_{i p}} \in\{$ true,false $\}$ is whether or not a shared IP is used in a transaction.

$\begin{aligned} \text { uncertain }_{\text {validaiton }}&=\left\{1 V_f=\text { null, }^{\text {is }} \text { common }_{\text {ip }}\right. \\ & =\text { false } 0 \text { otherwise } \\ & \end{aligned}$                (4)

Rule 3: Sensitive amount rule: The features that are related to amount are trade_amount, pay_single_limit and pay_accumulate_limit. For the aforementioned qualities, For the aforementioned qualities, fraudulent transactions typically contain the following attributes: To stay under the radar of a fraud-detection system and keep from being exposed as an impostor, a fraudster typically does a tiny test transaction before conducting a large transaction. This results in the addition of three characteristics to the raw data: sensitive_single_amount, sensitive_daily_amount and sensitive_test_amount. A={a₁, a₂, …, a_c} specifies the daily maximum that may be charged to a credit card, where a_c is the amount presently being tracked.

$sensitive _{\text {single }_{\text {amount }}}=\left\{1 a_c \in\left[A_1-\varepsilon_1, A_1\right] \quad 0 \quad otherwise\right.$              (5)

$sensitive_{\text {daily}_{\text {amount}}}=\left\{1 \quad \sum_{i=1}^c A a_i \in\left[A_2-\varepsilon_2, A_2\right] 0 ~otherwise\right.$             (6)

sensitive $_{\text {test }_{\text {amount }}}= \begin{cases}1 & a_c \in A_{\text {small }} 0 \text { otherwise }\end{cases}$           (7)

Rule 5: Elderly rule: Some fraud schemes target senior citizens in particular. Fraudsters will persuade elderly victims to transfer their entire account amount at once. According to the original feature trade_amount, card_balance. rule data now has an extra field labeled big_onetime_deal.

$big_{\text {onetime }_{\text {deal }}}=\{1 \quad$ trade_amount $\rightarrow$ card_balance 0 otherwise               (8)

Rule 8: Non-trusted time rule: Criminals that commit fraud often do it between the hours of 10:00 p.m. and 6:00 a.m. Based on the original feature white_list_mark and trade_time, the rule adds a new feature untrusted_time to raw data.

$\text {untrusted}_{ctime}=\{1$ white_list_mark $=$ $V_1($ not working $) 0~otherwise$        (9)

4.3 Selection of features using chimp optimization

The FS stage in applying machine learning algorithms is crucial. This is partly because the huge feature space in the dataset utilized for training and testing may have a detrimental effect on model performance. A researcher will choose one of the various FS methodologies depending on the kind of problem they are trying to solve. This work uses chimp optimization-based FS, which aids in choosing user features with suspect, fraudulent, and fraudulent results. The Ch_ORP is an example of swarm intelligence software that takes cues from chimpanzees' instinctive hunting behavior. The driver, barrier, chaser, and attacker are the four different types of agents in their neighborhood. In the optimization process, the "driver" guides the search, the "barrier" sets constraints, the "chaser" fine-tunes based on criteria, and the "attacker" disrupts by introducing variability. Together, they create a dynamic and adaptive optimization framework. Even if each chimp in a group has a distinct set of skills, these variations are necessary for accurate modeling of the hunting process. The mathematical description is as follows:

$X_{\text {chimp }}^{(t+1)}=X_{\text {prey }}^t-1\left|c \cdot X_{\text {prey }}^t-m \cdot X_{\text {chimp }}^t\right|$                (10)

where, $t$ is the current iteration, $X_{\text {prey }}^t$ and $X_{\text {chimp }}^t$ are the prey and chimp's position vectors, $a, m$, and $c$ are the coefficients stated as:

$a=2 \cdot f \cdot \operatorname{rand}_1-a$                             (11)

$c=2 \cdot$ rand $_2$                            (12)

$m=$ chaotic_value                            (13)

The m vector in these equations uses multiple chaotic vectors to indicate the sexual desire of the agents, and f is iteratively lowered nonlinearly in the interval [2.5, 0], whereas rand₁ and rand₂ are uniformly distributed between 0 and 1.

The stochastic population creation of agents kicks off the Ch_ORP. Like previous swarm-based techniques, this step requires a starting group of agents to be continuously developed across iterations. The four kinds of agents—driver, barrier, attacker, or chaser—are allocated at random. All groups will try to forecast the optimal prey placements; however, different methods will detail where the f vector is located and how the agents' positions are updated. The best position so far is that of the prey. Iterative adjustments to the c and m vectors enhance the avoidance of local minima and the speed of meeting. The assailant chimp often takes charge throughout the exploitation phase; however, the other agents may sometimes join in the actual hunting. Mathematical simulations of the hunting process are used because it is impossible to identify the prey's ideal position in advance. The other agents change their positions to accommodate the best agents, who are the pioneering chimpanzee drivers, attackers, chasers, and defenders. The position updating rule is represented by the equations below:

$x_1=x_{\text {attacker }}^t-a_1\left|c_1 x_{\text {attacker }}^t-m_1 x^t\right|$                            (14)

$x_2=x_{\text {harrier }}^t-a_2\left|c_2 x_{\text {barrier }}^t-m_2 x^t\right|$                            (15)

$x_3=x_{\text {chaser }}^t-a_3\left|c_3 x_{\text {chaser }}^t-m_3 x^t\right|$                            (16)

$x_4=x_{\text {driver }}^t-a_3\left|c_3 x_{\text {driver }}^t-m_4 x^t\right|$                            (17)

$x(t+1)=\frac{x_1+x_2+x_3+x_4}{4}$                            (18)

where, the m vector depicts the erratic actions taken by agents in the closing stages of a hunt to gather more meat, and the outcome is an increase in social favors such as grooming. In high-dimensional problems with many possible solutions, chaotic maps have been proven to aid in convergence and protect against the entanglement of local optima. Assume that half of the agents will act normally and the other half will update their position using chaotic methods. The theoretical definition of the updating formula for this method is:

$X_{\text {chimp }}^{(t+1)}=\left\{X_{\text {prey }}^t-\text { a.d } \text { if } \mu<\right.0.5$ chaotic$_{\text {value }}$ if $\geq 0.5$           (19)

where, 0.5 is a chance number between [0, 1]. As a consequence, the creation of random agents (possible solutions) serves as the first catalyst for Ch_ORP. Second, each agent is assigned to one of the previously mentioned four autonomous groups. Agents then use the provided classified technique to update their f vector. Then, with each cycle, the four groups assess the potential prey locations. The distances between the agents and the prey can then be updated. Finally, chaotic maps make it possible to accelerate convergence while avoiding local minima stagnation. The goal of this study is to zero in on the optimum collection of input characteristics that improves the model's accuracy without increasing its fitness.

$\operatorname{err}\left(x_i\right)=\text { actual }_{\text {output }\left(x_i\right)} -\text { model estimated }_{\text {output }\left(x_i\right)}$            (20)

$\text { fitness }(x)=\frac{\sum_{x=0}^n \operatorname{eerr}(x)}{n}$               (21)

Those whose fitness levels have been reversed and made higher than or equal to their original fitness levels are the ones who will take on the roles of the new people. The least physically fit members of society are chosen to reproduce. The machine then selects the best possible options.

Chimp Optimization-Based FS (Ch_ORP) differentiates with a bio-inspired technique that mimics chimp behavior. When compared to conventional approaches, genetic algorithms, machine learning models, ensemble methods, and hybrids, Ch_ORP provides unique flexibility and a decentralized swarm intelligence element, offering a significant edge in FS for fraud detection.

4.4 Algorithm for FS

Input: extracted features

Output: optimal selected features

$i n=\{\operatorname{ext}(1), \operatorname{ext}(2) \ldots . \operatorname{ext}(n)\}$

Update the community of chimps driver (dr), barrier (ba), chaser (ch), and attackers (att)

Initiate dr and ch

$X_{\text {chimp }}^{(t+1)}=X_{\text {prey }}^t-1\left|c \cdot X_{\text {prey }}^t-m \cdot X_{\text {chimp }}^t\right|$

Update the hunting process

Increase the number of agents at each iteration

Assign the agents to any one of category dr, ba, ch, att

Make these group to find the best prey (B)

$B=\left\{\right.$ best $_{\text {prey } 1}$, best $_{\text {prey } 2}, \ldots$ best $\left._{\text {preyn }}\right\}$

Initiate the exploration phase

Obtain the optimal position by hunting

Finalize the position of agents and prey and update the fitness value

The effectiveness of the suggested technique is examined in this part using three distinct datasets [19-21]. Tables 2-4 provide a description of the datasets.

Table 2. Credit card datasets

Table 3. Mortgage datasets

Table 4. Insurance datasets

Figure 3 and Table 5 illustrate the comparison of accuracy, MAE, and MSE, precision, recall and F1_Score for the credit card datasets.

The accuracy, MAE, MSE, precision, recall, and F1_Score of the ChOpt+LSTM model compared to the LSTM model presently used on the credit card data are all shown in Figure 3. The resulting figures provided that the ChOpt+LSTM model had enhanced accuracy of 99.18 percent, a lowered MAE of 25.7, a lower MSE of 16.3, precision of 98.54 percent, recall of 98.47 and F1_Score of 96.58 compared to the LSTM model's slightly decreased accuracy of 96.45 percent, increased MAE of 58.9, an increased MSE of 44.2, precision of 93.47 percent, recall of 92.04 and F1_Score of 90.09.

Accuracy, MAE, MSE, precision, recall, and F1 Score for identifying credit card fraud are compared in Figure 4 and Table 6. When compared to the LSTM model that is currently being applied to the Insurance Fraud data set, Figure 4 presents a comprehensive examination of the ChOpt+LSTM model's accuracy, MAE, MSE, precision, recall, and F1 Score. The ChOpt+LSTM model outperformed the LSTM model in every metric, with an improved accuracy of 99.51 percent, a lower MAE of 33.8, a lower MSE of 39.3, a higher recall of 95.88, and an F1 Score of 96.97 compared to the LSTM model's 97.00 percent, 48.5 percent, 42.0 percent, 92.71 percent, and 90.26 percent, respectively.

The accuracy, MAE and MSE, precision, recall, and F1 Score for identifying credit card fraud are compared in Figure 5 and Table 7, respectively. Figure 5 displays an in-depth analysis of the ChOpt+LSTM model's accuracy, MAE, MSE, precision, recall, and F1 Score in comparison to the LSTM model presently utilized on the mortgage sata. The ChOpt+LSTM model outperformed the LSTM model in every metric, including accuracy (99.34%), MAE (26.8), MSE (12.0%), recall (99.95%), and F1 Score (97.66%), compared to the LSTM model's (96.99%) performance in these areas.

Tables 8-10 show the accuracy for various datasets. This table shows that the proposed approach outperforms the state-of-the-art alternatives with respect to accuracy.

4a.jpg

4b.jpg

4c.jpg

Figure 4. Comparative analysis of insurance datasets

Hence, the Chimp Optimization-Based FS (ChOpt) provides targeted and adaptable FS, which accounts for the ChOpt+LSTM model's excellent performance. The swarm intelligence-inspired design of ChOpt enables dynamic exploration of feature combinations, identifying complex fraud flags. The synergy with LSTM exploits strengths in FS and capturing temporal relationships, making ChOpt+LSTM useful in datasets with complicated fraud patterns.

Table 8. Comparative analysis of various methods for credit card datasets

References	Techniques	Accuracy (%)
Developed method	ChOpt+LSTM	99.18
Shahapurkar [22]	AFD-HHM	96.00
Hajek et al. [23]	GBDT+XGBoost	52.45
Arun et al. [24]	BEPO-OGRU	94.74

Table 9. Comparative analysis of various methods for insurance datasets

References	Techniques	Accuracy (%)
Developed method	ChOpt+LSTM	99.51
Parnian et al. [25]	SOA+kNN	98.08
Panda et al. [26]	PKRR+ISSO	98.5
Bharat et al. [27]	RHOFS+kNN	99.61

Table 10. Comparative analysis of various methods for mortgage datasets

References	Techniques	Accuracy (%)
Developed method	ChOpt+LSTM	99.34
Ojugo and Nwankwo [28]	GA+MNN	74
Keswani et al. [29]	LR+SMOTE	97
Bharat et al. [27]	RHOFS+NB	98.85

5a.jpg

5b.jpg

5c.jpg

Figure 5. Comparative analysis of mortgage datasets

8.1 Comparative analysis with existing techniques

Tables 11-13 show the MAE and MSE values for various datasets. According to that table, the suggested approach has fewer mistakes than several other existing methods.

Tables 14-16 show the precision, recall, and F1Score values for various datasets. The values in the suggested technique are higher in that table when compared to the various existing methods.