Enhanced Detection of Text and Image Spam Using Cost-Sensitive Deep Learning

The internet has become an integral part of human existence, with over 4.5 billion people using it for daily activities. Most internet users consider email a reliable communication tool [1], and its effectiveness has grown over time. However, as email usage increases, so does the number of spam attacks. Spammers can transmit spam from anywhere with internet access, as long as they have the receiver's data and are not solicited. Spam emails are sent without the receiver's data [2]. Phishing emails, often containing fake content or links to malicious websites [3], are widely used to collect users' sensitive information for their own financial gain without their consent. Despite advancements in spam filtering software, there is no reliable mechanism to distinguish genuine and harmful emails due to the ever-changing nature of spam content, and despite anti-spam tools, naive end-users continue to fall prey to this harmful trap [4]. Spam has been sent for three or four decades, and, despite this, it continues to be sent. The Personalized Email Prioritization (PEP) issue is a key impediment, with the goal of assisting users in rating the relevancy of email communications [5, 6]. This is yet another issue with email classification; one potential solution would be to assign low priority to suspicious emails [7-9]. Spam emails are those that were sent without the receiver's data and were not solicited. Machine learning is a powerful tool for detecting image spam shown in Figure 1, but it's crucial to recognize its limitations. Machine learning models can be vulnerable to biased training data, which can make it difficult to detect new spam types [10]. Adversaries can exploit this vulnerability by creating intentionally obfuscated or manipulated images, which can bypass the model's detection mechanisms, particularly for image spam relying on visual elements like logos, watermarks, or text. Intelligent detection is crucial for identifying changes in spam email content and individual user priorities [11, 12]. Deep learning is a powerful technique that has shown great potential in various domains, including image processing and pattern recognition. It involves training deep neural networks with multiple layers to learn representations and features from data automatically. With image spam detection, deep learning can be utilized to extract relevant features from images that distinguish between spam and non-spam. By leveraging multiple layers of neurons, deep learning models can learn complex patterns and relationships in the data, enhancing the accuracy of the classifier. Supervised learning approaches can be used to train deep learning models for image spam detection using a labeled dataset where each image is annotated as spam or non-spam. CNNs [13] are a common approach to deep learning for image classification tasks. They are particularly effective for image analysis because of their ability to capture spatial hierarchies and local patterns. Unsupervised feature learning algorithms, such as autoencoders or generative adversarial networks (GANs), can be employed to learn meaningful representations from the image data, which can then be used for spam detection. In conclusion, deep learning is a powerful tool for image spam identification because of its ability to automatically learn complex features and patterns from image data, leading to more accurate and robust classifiers. CNNs are built from repeated bits of information organized in this way and have been successfully employed in image spam detection [14].

1.png

Figure 1. Examples of spam images

The data augmentation [15] approach generates new training dataset samples by making random changes to existing datasets. This has a number of consequences, including faster learning process convergence and less overfitting [16]. Starting from scratch with a deep neural network demands a significant amount of data and computing resources. Transfer learning has become an essential tool for building efficient text and image spam classification models. It enables faster training, higher levels of efficiency, and greater adaptation to emerging spam threats by using pre-trained models and fine-tuning them for particular tasks [17]. To do this, most of the layers' parameters are inaccessible, and the learning process only adjusts a subset of the layer's parameters. It is also useful for dealing with data scarcity, assuming that the pre-trained model was trained with a sufficient amount of data. They differentiated between images classified as spam and those labeled ham. The Dredge dataset [18], the Image Spam Hunter (ISH) dataset [19], and the enhanced dataset [20] are some of the common datasets used in conjunction with a variety of deep learning models and their respective applications.

The objective and motivation behind the research are mentioned below:

• Our paper introduces a novel approach that integrates both text and image data to enhance the accuracy of spam image classification.
• We propose an innovative technique that automatically calculates class-dependent costs based on data statistics, such as distribution and reparability measurements.
• Unlike existing methods, our approach applies class-dependent costs exclusively during the training phase. This allows for predictions without the need to alter the trained network post-learning, streamlining the inference process while maintaining the integrity of the trained CNN parameters.

In the following section, we discussed the related work in our field. The third section explains the proposed work. Sections 4 and 5 covered the results and conclusion.

The method employs text and image spam filtering techniques to enhance accuracy. The model architecture shown in Figure 2 includes data preprocessing, feature selection, and classification models. Data normalization and standardization are performed in stage 1, followed by feature extraction selection in stage 2. In stage 3, various classification models are trained and tested for text and image features. Stage 4 evaluates performance metrics, while stage 5 demonstrates spam or ham detection.

3.1 Datasets description

In this paper, we use two datasets, namely the Spambase and ISH datasets.

Spambase dataset [27]: The dataset contains 4,601 emails, which are categorized as spam or not. The dataset comprises 57 numerical features that indicate email features, including word and character rate, capital letter sequence length, and symbol occurrence. The Spambase dataset is used for spam filtering and machine learning classification. The dataset is difficult to deal with since no one feature separates spam from legitimate emails.

ISH: In this paper, we used ISH [28], which was the first dataset made available to the public. As seen in Figure 3, it contains two unique sets of genuine images preserved in the JPEG file format. Figure 3a shows some images from 810 ham images. Similarly, the 928 spam images were collected from real spam e-mails (see Figure 3b). It should be noted that it corrupted eight spam images during the extraction process from the spam image collection and is thus not included in the sample set. Table 1 lists all the details of an ISH dataset, which contains 1730 images.

Table 1. Summary of ISH dataset

3.2 Data preprocessing

Preprocessing organizes and modifies raw data before training and evaluating classifiers. A data mining approach organizes raw data into usable formats. Preprocessing is the initial step in building a machine-learning model. This stage takes real-world data, which often contains errors, inconsistencies, and inaccuracies, and converts it into accurate, precise, and usable input variables and patterns.

2.png

Figure 2. Architecture of proposed model

3.png

Figure 3. Some examples of images from ISH dataset

3.3 Data extraction

Feature extraction simplifies large raw datasets, which depend on the initial dataset. At this stage of the process, we can extract data from the dataset, such as variables, attributes, or classes. One of the most critical processes in training the model is feature extraction [29]. This contributes to more reliable and accurate outcomes. We refer to the strategy of selecting a few significant variables that accurately define the data throughout the feature extraction process as feature selection. The method employed is feature selection from among the various attributes. Following that, we construct the model by combining the traits or variables that were chosen. If we follow the feature selection procedure correctly, it will minimize the time spent on the model.

Monarch Butterfly Optimization: After discussing the concept of a self-adaptive strategy, an enhanced MBO algorithm called monarch butterfly optimization with a self-adaptive population (SPMBO) is presented [30]. New individuals in SPMBO are only accepted by those who are superior to previous generations.

3.4 Proposed model

The primary goal of the proposed model is to build a deep learning-based model and evaluate and compare it with several machine learning algorithms from more traditional approaches that were tested.

Gaussian Naive Bayes: Supervised machine learning has been used to detect spam. The ability to discriminate between various things based on defined features is a key component of its success. We calculated the probability of a word or event reoccurring in the future using this method [31]. For instance, if an e-mail has a word that is only present in spam e-mails and not in ham e-mails, the algorithm will almost certainly classify it as spam.

$(c / x)=(P(x / c) P(c) /(P(x)))$                       (1)

$P(x)=\sum_y P(x / c P(c))$                     (2)

In this case, x represents function vectors, and c denotes a class variable.

SVM: Another tool for supervised machine learning is the SVM. It will only function with pre-classified datasets. When training SVMs, it is most often used for classification and regression models. The data classification is more reliable than another model when there is a limited amount of labeled data. The SVM serves better for the classification. To distinguish between positive and negative values in the dataset (also known as spam and ham), a hyperplane is used. Then, determine which values are close to the decision boundary. Figure 4 is an illustration of the SVM.

4.png

Figure 4. SVM [32]

KNN: This serves as a classification method that represents items as points in a space and calculates the distances between them. In the learning phase, the algorithm assigns training data points to clusters based on their proximity to the center. The algorithm takes a parameter, k, which represents the number of nearest neighbors to consider, and its value can be optimized. The choice of k has an impact on the accuracy of the classification. The k nearest neighbors are the instances in the training sample that are closest to the object being classified. The object is assigned to the class that shares the most attributes with its k nearest neighbors. However, the KNN algorithm can be unstable when dealing with outliers and does not perform well with many features.

Logistic Regression: This is a suitable method for modeling and explaining the relationship between a binary response variable and instructive modules [33]. It is used to analyze data and predict the probability of assigning a specific class, where the values range from 0 to 1. Logistic regression provides a way to model the likelihood of an event occurring based on the given inputs.

Decision tree: This is a hierarchically structured structure that divides the feature space into subspaces. Then, for each object encompassed within this subspace, predict the result, in which algorithm overfitting is a typical issue.

Random forest: This prediction method relies on the construction of trees. By combining multiple trees into a forest, the predictive power of each tree can be enhanced. In the training phase, we construct multiple decision trees based on the programmer's specifications [34]. These trees are then utilized to predict the class. The prediction is made by examining the class votes from each tree, and the output is determined by selecting the class with the highest number of votes.

Adaptive Boosting: AdaBoost is a machine learning ensemble method used for classification tasks. It involves iteratively training weak learners, such as decision trees, into a single strong learner with improved accuracy. The process involves assigning equal weights to all data points, evaluating the learner's performance, and increasing or decreasing the weights of misclassified data points.

Cost-insensitive models: Cost-insensitive models refer to machine learning models that do not explicitly consider the cost associated with misclassifying instances during the training process. In many classification problems, especially in scenarios where the cost of false positives and false negatives is uneven, it becomes important to account for these costs to optimize the model's performance.

In a cost-insensitive model, the misclassifications are treated equally during the training process, and the model is optimized based on a standard loss function without considering the specific costs associated with different types of errors. However, there are scenarios where cost-insensitive models might be appropriate. If the cost associated with misclassifications is relatively uniform across different classes, or if the focus is solely on overall accuracy without considering the consequences of specific errors, a cost-insensitive approach may be suitable.

Cost-sensitive models: Cost-sensitive models are machine learning models that explicitly consider the costs associated with different types of errors during the training process. In many real-world scenarios, the consequences of false positives and false negatives can vary, and it's important to build models that prioritize minimizing the total cost rather than simply optimizing for overall accuracy.

3.5 CNN model

Figure 5 shows the CNN 2D image processing architecture. An ordinary CNN classifies the visible spam and ham data. CNNs are best suited for text and image spam classification. CNNs mimic the way the human brain interprets images. By not integrating all the nodes, the approach reduces the processing time required and increases efficiency. Also, it successfully discriminates and emphasizes data as features with nearby images with the help of spatial data. This facilitates comparison. CNN models' components are in charge of extracting and classifying image features.

5.png

Figure 5. Illustration of the architecture of CNN 2D

3.6 Proposed cost matrix

We denote the proposed cost matrix with the symbol θ, which is appropriate for training the input data for feature extraction. However, as shown in Figure 6, it can be used to alter the output of a CNN's last layer for activation of layers compressed amid 0 and 1 in advance, determining the loss of the model.

6.png

Figure 6. Simplified overview of CNN parameters

The CNN classifies data based on the highest latent score during the classification procedure. During training, the classifier weights are adjusted to alter the confidences or probabilities of the classifier. This is done to ensure that the targeted class receives the highest latent score while the other sequences have significantly lower scores. To enhance classification, "score-level costs" are added to the trainset, affecting the CNN outputs (o) through a mathematically indicated cost matrix (ξ).

$y^{(i)}=\mathcal{F}\left(\xi_p, o^{(i)}\right), \quad ; y_p^{(i)} \geq y_j^{(i)}, \forall j \neq p$                                (3)

where, y represents the adjusted output, and p indicates the desired class. The confidence of a classifier is significantly influenced by its “score-level costs.” This type of perturbation enables the classifier to prioritize “less common” and “challenging-to-separate classes.”

3.7 Cost-sensitive surrogate losses

This CNN training method overcomes class imbalance. To achieve this, we propose a cost-sensitive error function, the mean loss across the training set:

$E((\theta, \xi))=\frac{1}{M} \sum_{i=1}^M \ell\left(d^{(i)}, y_{\theta, \xi}^{(i)}\right)$                              (4)

where, ξ denotes class-sensitive costs, M and N denote is the numbers of training features and output layer neurons, y designates the predicted output, and $d \in\{0,1\}^{1 \times N}$ denotes the desired output when $\sum_n d_n=1$. We will describe a single data instance and not directly address y dependence on parameters (θ; ξ). The optimization goal is to find ideal parameters (θ∗; ξ∗) for the lowest cost E∗, as the error increases as the prediction system performs catastrophically in training.

$\left(\boldsymbol{\theta}^*, \boldsymbol{\xi}^*\right)=\arg \min _{\theta, \xi} E(\theta, \boldsymbol{\xi})$                              (5)

In Eq. (4), the loss function $\ell($.) is any variant of the three Cost-Sensitive (CS) losses:

(i) This MSE loss minimizes the squared error between expected output and ground-truth.

$l(d, y)=\frac{1}{2} \sum_n\left(d_n-y_n\right)^2$                                  (6)

where, y_n represents the previous layer output:

$y_n=\frac{1}{1+\exp \left(-o_n \xi_{p, n}\right)}$                                  (7)

(ii) The hinge loss function exploits margins among classes, expressed as:

$l(d, y)=-\sum_n \max (0,1)-\left(2 d_n-1\right) y_n$                                  (8)

where, y_n is the “previous layer output” and ξ is the cost.

$y_n=o_n \xi_{p, n}$                                  (9)

(iii) The CE loss function exploits the prediction's accuracy, as shown by:

$l(d, y)=-\sum_n\left(d_n \log y_n\right)$                                  (10)

where, y_n includes the “class-dependent cost” (ξ) and is associated to the output through the “soft-max function”.

$y_n=\frac{\xi_{p, n} \exp \left(o_n\right)}{\sum_k \xi_{p, k} \exp \left(o_n\right)}$                                  (11)

3.8 Cost-sensitive classifier

The ability of cost-sensitive classifiers to distinguish between treatments for majority and minority classes is their most significant benefit. They also assess the financial costs of misclassification. As stated in the study [35], the negative (majority) and positive (minority) classes are represented by the values 0 and 1, respectively. The rows represent the current classes, while the columns represent the predicted classes. The letters TP and TN indicate correctly classified samples as positive and negative, while FP and FN indicate incorrectly classified samples as positive and negative. From the confusion matrix, the accuracy and geometric mean were determined as follows: Eqs. (12) and (13).

$G_{\text {mean }}=\sqrt{\frac{T P}{T P+F N} \times \frac{F P}{T N+F P}}$                                  (12)

$A c c=\frac{T P+T N}{T P+T N+F P+F N}$                                  (13)

The estimated risk is shown in Eq. (14) according to the least expected cost principle.

$R(i \mid S)=\sum_j P(j \mid S) C(j, i)$                                  (14)

In Eq. (14), $R(i \mid S)$ denotes estimated risk of classifying for the posterior probability is $P(j \mid S)$ and the misclassification cost is denoted as $C(j, i)$.

Also, posterior probability calculation is always challenging. Thus, empirical risk in mathematical form, such as Eqs. (15) and (16).

$\hat{R}_{\boldsymbol{l}}(\boldsymbol{o})=\boldsymbol{E}_{S, Y}[\boldsymbol{l}]=\frac{1}{n} \sum_{i=1}^n l\left(C, d^{(i)}, o^{(i)}\right)$                                  (15)

$C=\left\{\begin{array}{c}C_{p, q}=1, p=q \\ C_{p, q}=I R, p \neq q\end{array}\right.$                                  (16)

The relevant class labels are represented by Y. The total number of multivariate time series occurrences is given by n. When the projected class q matches the actual class p, the cost will be set as the imbalance ratio. The network's loss function is represented by l ().

3.9 Proposed cost-sensitive learning strategy

Cost-sensitive models are useful for text and image spam classification by accounting for the specific costs associated with misclassifications. They help address the imbalance between false positives and false negatives in spam classification. By customizing the loss function to reflect these costs, the model can minimize overall costs rather than focusing solely on accuracy. Cost-sensitive learning allows the model to adapt to new types of spam by considering the specific costs of misclassifying unseen instances. Adjusting the classification threshold based on the trade-off between false positives and false negatives can help minimize the cost of misclassifications in spam classification.

If a penalty for cost-sensitive misclassification entails using an imbalanced ratio, this may help alleviate the overall class imbalance problem. Nonetheless, the fixed cost matrix could not account for the uneven dissemination such as trained CNN data. As a result, the proposed system used a misclassification cost weight that could be updated iteratively and changed dynamically. This was constructed for uneven dissemination of both trained data and minibatches. The previously stated convolutional classifiers were modified, utilizing the cost-sensitive learning technique to deal with ITSC concerns.

The nth training instance's cross-entropy loss could be expressed as follows:

$\begin{aligned} \operatorname{LOSS}(\theta)=\lambda \times d_n & \times\left(-\ln \left(y_n\right)\right. \left.+\left(1-d_n\right) \times\left(-\ln \left(1-y_n\right)\right)\right)\end{aligned}$                                  (17)

where, $\theta$ denotes weight parameters.

The cost of misclassification is assigned a cost value. The desired output, denoted by d_n, and the predicted output, denoted by y_n, are both listed in the nth training instance. The global loss optimization is represented by Eqs. (18)-(20).

$\begin{aligned} & E(\theta)=\frac{1}{n^{p o s}} \sum_{i=1}^{n^{p o s}} \operatorname{LOSS} S^{p o s}\left(\theta^{p o s}, \lambda_n^{p o s}\right) +\frac{1}{n^{n e g}} \sum_{i=1}^{n^{p o s}} \operatorname{LOSS}{ }^{n e g}\left(\theta^{\text {neg }}, \lambda_n^{\text {neg }}\right) \\ & \end{aligned}$                                  (18)

${{\lambda }_{n}}=\left\{ \begin{matrix}   I{{R}^{normal}}\times exp\left( -\frac{G_{mean}^{batch}}{2} \right)  \\   \times exp\left( \frac{Ac{{c}^{batch}}}{2} \right)  \\   1,~if~n\epsilon ~pos~  \\\end{matrix} \right.,~if~n\epsilon neg$        (19)

$\begin{aligned} & \left(\theta^*\right)=\operatorname{argmin} E(\theta) \end{aligned}$                                  (20)

$I R^{\text {normal }}$ denotes the overall imbalance ratio. The local metrics $G_{\text {mean }}^{\text {batch }}$ and $A c c^{\text {batch }}$ are updated after each minibatch. For input training unbalanced temporal sequence sets, a random shuffle method was utilized to allocate minibatches from scrambled time series data. This strategy avoided minibatch deficits in minority data and improved classifier generalization.