Hybrid Deep Learning Framework via Early Feature Fusion for XSS Attacks Detection

The extensive use of the Internet, especially web applications and services, has become an integral part of daily life due to the development of dynamic websites that have simplified users' interaction processes. Unfortunately, these websites have become a threat or vulnerable, which have a significant impact on online security and allows attackers to target internet-connected devices [1].

With these potential threats, cybersecurity has become vital on a wide scale in social, industrial and economic contexts, as developers and cybersecurity professionals recognize the possibility of trained attackers launching complex and mysterious attacks [2, 3].

Cross-site scripting (XSS) attacks are one of the most methods leveraged by attackers to inject malicious scripts into legitimate websites to steal information, gain unauthorized access, or perform other illegal activities [4].

 Although there are comprehensive strategies for mitigating XSS attacks (which have included integrating robust security controls, regular security audits, and classic detection methods such as signature-based detection), skilled or sophisticated attackers often evade detection, compromising the security of online applications [5]; where payloads containing specialized code or deceptive HTML tags can bypass traditional filters.

Consequently, the evolution in artificial intelligence and machine learning enhances online application security that can promptly identify dynamically evolving attack patterns [6, 7]. Based on vast amounts of data, these technologies offer a strong and proactive security system that can recognize hidden patterns and criminal strategies [8].

Unlike traditional detection frameworks, ML- and DL-based systems can learn from the evolving behavior of dynamic threats and continuously improve their detection performance [9].

The methods of deep learning of the semantic and structural representation of XSS payloads have demonstrated efficient performance through the essential metric results [10]; the detection rate of XSS attacks has significantly increased due to the application of these techniques [11]. Combining machine learning and deep learning techniques, which use artificial intelligence to scan web content, find patterns, and search for potentially harmful code [12], such as malicious JavaScript code or input validation issues, has been shown to achieve better detection of XSS attacks [13].

Through constant learning of the emerging attack trends and evolving threat patterns, AI-powered XSS detection systems can enhance proactive defense mechanisms [14], thereby strengthening defenses against the common and emerging threats [15].

Despite significant progress in applying deep learning and machine learning techniques for XSS attack detection, many methods proposed so far continue to process symbolic features, such as Term Frequency-Inverse Document Frequency (TF-IDF) and sequential forms, such as token embeddings, separately. Separate processing of these prevents models from fully benefiting from the promised synergies of combining frequency-based statistical features with sequential contextual patterns. Previous works have rarely addressed early fusion methods that merge the two complementary perspectives, leaving a significant research gap and limiting the development of more robust and general detection systems.

In addition to these limitations, several XSS payloads used in modern attacks incorporate varying layers of structural manipulation, encoding, and context-dependent behavior, which makes it increasingly difficult for models relying on a single feature perspective to generalize across different obfuscation styles. This highlights the need for detection approaches that can capture both statistical and contextual cues in a unified manner.

Addressing these features separately limits the model's ability to detect payloads that combine structural manipulation with subtle character-level signals, which is a common approach in modern obfuscation techniques. The effectiveness of current fusion techniques to capture joint patterns suffers since they usually combine information after each branch has learned independently. By allowing statistical and sequential representations to be trained simultaneously from the input stage, Early Feature Fusion directly closes this gap and enables the model to recognize complementing inputs that each representation can capture on its own.

This research paper proposes an early fusion strategy, that constantly combines contextual and statistical features to make models to comprehend the syntactic and semantic nature of XSS attacks and categorize them more precisely and reliably.

The primary contributions of this research toward XSS attack detection are as follows:

• Feature Fusion Architecture: A preliminary fusion method is presented that combines statistical features obtained through TF-IDF with temporal representations that are learned by deep recurrent networks (LSTM or GRU); the fusions obtain both syntactic frequency-based patterns and contextual dependencies of XSS payloads.

• Reducing Dimensionality Using Principal Component Analysis (PCA): In high-dimensional statistical representations, for example, those based on TF-IDF, the multiplicity of dimensions can affect model generalization and potentially increase computational costs. Therefore, employing PCA as an approach to reducing the dimensions for common and redundant features enables the model to reduce computational costs, increase learning efficiency, and decrease the likelihood of overfitting, especially in fusion-based architectures.

• Performance Assessment: Multiple variants of the suggested architecture are trained and tested. Performance is assessed using a variety of classification metrics, including accuracy, precision, recall, and F1-score.

The rest of the research comes under Section 2, is dependent upon the investigation of comparative literature. Section 3 describes our proposed research approach. Findings from the method used in Section 4. Discussion in section 5. Finally, the conclusion in Section 6.

3.6 Dimensionality reduction (PCA)

To address the potential high dimensionality of TF-IDF vectors, PCA is applied prior to fusion in two of the models (PCA-LSTM and PCA-GRU). The number of components is empirically set to 100. PCA serves to reduce noise, enhance generalization, and accelerate convergence during training.

3.7 Model architecture and classifier

The proposed models consist of two input branches: one for padded sequences passed through an embedding layer and recurrent block (LSTM or GRU), and another for TF-IDF. The branches are merged and followed by dense layers with dropout regularization. The final output layer uses a sigmoid activation function for binary classification. Table 2 illustrates the hyperparameters configurations of the proposed model.

3.8 Training and hyperparameter settings

All models were trained for a maximum of 50 epochs with a batch size of 32, using the Adam optimizer and binary cross-entropy loss function. A validation split of 20% was applied to monitor model performance during training. An early stopping mechanism was implemented to prevent overfitting and optimize training efficiency. The embedding layer produced 100-dimensional vectors, and the TF-IDF word vectorizer used a dimensionality of 1,000. Both fully connected layers included a dropout rate of 0.3 to mitigate overfitting. No hyperparameter tuning or optimization search was performed; instead, the same fixed parameters were applied across all model variations to ensure a fair comparison and to isolate the effect of early fusion and dimensionality reduction techniques.

3.9 Evaluation metrics

Performance was evaluated using an eight-point, well-thought-through, well-referenced, and well-established measure set that can be used in the case of binary classification problems. Out of these, Accuracy measures overall correctness in predictions; Precision denotes the value of true positives over all predicted positives; and Recall measures the ability to classify true malware instances rightly. To trade off Precision and Recall into one harmonic value, the term F1-score was utilized. To calculate credibility in predictions and statistical significance, the Matthews Correlation Coefficient (MCC) was employed. To measure the proper classification of benign samples, the terms 'specificity' (True Negative Rate) and (False Negative Rate) were utilized, while ROC–AUC (Receiver Operating Characteristic – Area Under the Curve) is a metric that evaluates the model’s ability to distinguish between the positive and negative classes across all possible decision thresholds. The ROC curve is plotted using the True Positive Rate (TPR) against the False Positive Rate (FPR); this makes it particularly useful in evaluating how well the model separates malicious XSS payloads from legitimate requests in a balanced way. The same measure set was employed for experiments to calculate and achieve a multidimensional and interpretable process of assessment. All these measures were calculated based on the equations illustrated below Eqs. (1)-(8), which are the mathematical expressions for Accuracy, Precision, Recall, F1-score, MCC, FNR, TNR, and ROC–AUC, respectively.

Accuracy $=\frac{T P+T N}{T P+T N+F P+F N}$         (1)

Precision $=\frac{T P}{T P+F P}$           (2)

Recall $=\frac{T P}{T P+F N}$          (3)

$F 1-$ score $=2 \times \frac{Precision × Recall}{Precision+Recall}$        (4)

$M C C=\frac{\operatorname{cov}(X, Y)}{\sqrt{(\operatorname{cov}(X, X) \times \operatorname{cov}(Y, Y))}}$            (5)

$F N R=\frac{F N}{F N+T P}$            (6)

$T N R=\frac{T N}{T N+F P}$            (7)

$\begin{aligned} R O C-A U C= & \sum_{i=1}^n\left(\frac{T P R_i+T P R_{i-1}}{2}\right) \times\left(F P R_i\right. \left.-F P R_{i-1}\right)\end{aligned}$       (8)

Our proposed system's findings demonstrated that using an early feature extraction fusion strategy within TF-IDF-extracted statistical representations and sequential contextual representations generated by LSTM or GRU models allowed deep learning techniques to understand the syntactic and semantic patterns of XSS attacks in a unified, integrated phase. This strategy created a learning environment that allowed the model to detect specific correlations between features from the beginning, as opposed to late fusion, which frequently examines features individually and then combines their results later, leading in the loss of key interaction information. Applying PCA to the TF-IDF outcomes before fusion also helped to reduce dimensionality and remove redundancy, which improved the models' capability to generalize and accelerate the training process while maintaining accuracy. This was reflected in the convergent performance by the four proposed models' high and convergent performance, with the GRU + TFIDF model surpassing the others on the majority of criteria. Figure 14 shows that the proposed model (GRU + TFIDF) achieved better results in terms of detection capabilities compared with the previous researches (mentioned in the related works) evaluated with the same dataset, demonstrating the efficiency of the methods used.

Figure 14. The related work based on Dataset_1

While the results are promising, claims of robustness and scalability remain limited to the two datasets evaluated and do not yet account for hostile or realistic deployment scenarios. Additional difficulties are introduced by practical integration into current Web Application Firewalls (WAFs), such as inference-time delay, resource constraints, and the requirement for regular model upgrades to support novel obfuscation techniques. While heavily obfuscated or borderline benign samples continue to be more challenging, the observed misclassification patterns indicate that early fusion offers special benefits for payloads combining statistical and contextual factors, such as partially obfuscated JavaScript fragments or encoded injections. All these elements highlight the current framework's practical limits and point to crucial areas for future research.