Use of Multiple Inputs and a Hybrid Deep Learning Model for Verifying the Authenticity of Social Media Posts ^†

Abstract

With the rise of social media platforms and the vast amount of text content generated on these platforms, text data forensics has emerged as a new area of research that aims to verify posts’ authenticity by analyzing textual content. This study proposes an innovative hybrid framework for detecting fake content on social media by examining both the text and metadata of Twitter posts. The metadata are fed into a feature selection method to select the most beneficial features. Using multiple inputs, a hybrid deep learning framework is proposed to classify Twitter posts as real or fake, where fake content is defined as posts containing misleading information. This research significantly contributes to the field of text data forensics by enhancing the detection of such fake texts. A recent comprehensive dataset for text data forensics called CIC Truth Seeker Dataset 2023 was used to assess the effectiveness of the proposed framework; the proposed framework uses long short-term memory (LSTM) to process textual data and hybrid residual neural network (ResNet) and deep neural network (DNN) layers for metadata. The framework has shown promising results during its preliminary evaluations. The paper examines the proposed model’s architecture and performance while highlighting potential improvements in privacy, ethics, real-time deployment, and implementation limitations to emphasize its broader impact.

1. Introduction

Digital investigation involves identifying, collecting, and acquiring digital evidence to examine, analyze, and draw a valid conclusion regarding a specific violation [1]. Social media platforms have become integral to communication and data sharing, generating vast amounts of data that can be challenging to analyze using traditional digital investigation techniques [2]. Hence, performing conventional digital investigations, which are carried out manually by forensic experts, can be an overwhelming task [3]. Intelligent technologies such as machine learning (ML) and deep learning (DL) are crucial in bridging the gap and aiding digital forensics by automating the analysis of the vast amounts of data generated by social media platforms [4]. Text data forensics, a rapidly evolving sub-discipline, leverages ML to investigate social media data to identify misleading content. The objective is to analyze textual information for identifying potential fake content.

Text data forensics is a field that analyzes large amounts of text data to identify patterns, irregularities, anomalies, and digital evidence that could indicate misleading content. It benefits various governmental sectors, law enforcement, cybersecurity officials, and corporate brand monitoring. Identifying potential threats and analyzing market trends are two practical uses of text data forensics. Additionally, integrating text mining with social network analysis enhances investigations, enabling a more efficient correlation of user behavior and textual patterns. Machine learning can classify text and metadata as fake or real to identify misleading information. However, a machine learning-based online system used for verifying fake content in digital forensics must consider the sensitivity of its findings, as misclassifications could have serious legal and security implications. Additionally, the effectiveness of machine learning models, particularly deep learning models, depends heavily on the quality and balance of the available data [5,6].

Unstructured or imbalanced data can lead to decreased model accuracy and real-world applicability. Addressing class imbalance and noisy data is crucial in minimizing false positives and negatives, which can undermine trust in automated forensic systems. Many studies in social media forensics have primarily focused on analyzing text or simple metadata features. In contrast, this study combines text and metadata using a hybrid architecture that includes LSTM, ResNet, and DNN components. This multi-input design enables us to capture subtle user-level signals and the semantic context within the text. This approach sets our work apart from previous research and has the potential to enhance interpretability in digital forensics applications.

This article proposes a deep learning framework for text data forensics, specifically examining Twitter data to verify the authenticity of social media posts. Two separate feeds of input—namely, tweets and metadata—are preprocessed and later combined into a single model for enhanced performance. The metadata are first fed into a feature selection method, and three features are generated from the tweets and added to the metadata. This framework uses the random forest ensemble method to select important metadata features. The text is input into a hybrid deep learning framework along with its metadata. The data are processed using a combination of LSTM for the textual data and ResNet and DNN layers for the metadata. This multimodal input strategy is anticipated to produce more reliable results than methods that depend solely on text.

This article extends our previously published conference paper [7]. The main contributions of this extended paper are summarized as follows: we expand the evaluation metrics beyond accuracy to include precision, recall, F1-score, and AUC-ROC, ensuring a more robust assessment of model performance in potentially imbalanced datasets. We directly compare our ResNet–DNN metadata approach and TabTransformer, highlighting how attention-based architectures for tabular data compare with our lightweight residual blocks. We incorporate hyperparameter tuning to refine key aspects of our hybrid deep learning architecture, including LSTM units, dropout rates, and optimizers, leading to improved performance. We analyze practical constraints like computer demands, privacy issues, and problems with deploying solutions in real time on large social media platforms. This understanding shows the strengths and weaknesses of our framework for detecting fake posts in social media forensics.

The rest of the paper is structured as follows: Section 2 presents the related work on identifying potentially fake social media posts, emphasizing different deep learning and feature engineering approaches. At the end of Section 2, we also concisely compare our results with existing methods. Section 3 outlines the CIC Truth Seeker Dataset 2023, detailing key preprocessing steps and data division strategies. Section 4 introduces the proposed multiple inputs and hybrid deep learning framework. Section 5 demonstrates the experiments and discusses the results, along with potential avenues for further optimization. Section 6 summarizes the key contributions and discusses the method’s limitations, including data availability, privacy concerns, and real-world scalability. It also suggests future research directions, such as applying the framework to platforms like Facebook and Reddit and expanding it to multilingual contexts.

2. Related Work

Sharrab et al. [2] proposed a DNN-based approach to classify tweets as suspicious or normal. The authors utilized LSTM’s analytical and identification capabilities to enhance text data forensics in social media. Based on the preliminary experiments, it was shown that the proposed approach was highly effective and achieved very promising results. The authors plan to explore various applications, such as detecting misinformation, identifying criminal activity, and preventing online harassment. However, the authors rely on the text and ignore the metadata, which might help to improve the performance and broader interpretability of the model.

Kavin et al. [8] compared the effectiveness of three machine learning algorithms in detecting fake and real social media accounts. The investigated classifiers include support vector machines (SVMs), the random forest ensemble method, and an artificial neural network. The results indicate that the SVM outperforms the other two methods regarding recall, precision, and F-measure. The authors suggest using an SVM model to detect inappropriate content, highlighting the value of simpler ML methods when data quality or volume is limited.

Abarna et al. [9] conducted a lexical analysis of text using a traditional scheme and used a fast text model to improve the word ordering, thereby enhancing computational efficiency. The authors employed various techniques to extract features and analyze text intention, leveraging word frequency and bully–victim participation scores. The proposed model outperformed traditional machine learning algorithms, showing that simple lexical approaches can produce robust results in specific contexts.

Borkar et al. [10] introduced a method that uses clustering and classification techniques to identify fraud. To preprocess social media data, they used natural language processing (NLP) techniques, including vectorization, data normalization, and dimensionality reduction. A behavioral analysis technique was further used to extract meaningful characteristics from each profile. Clustering approaches were then employed to differentiate between real and fake profiles, and a recurrent neural network classified social media profiles accordingly. The results demonstrated strong applicability in real-world social media contexts, highlighting the significance of multi-step feature extraction processes.

Ahmad et al. [11] proposed an approach for identifying terrorism-related content. The authors analyzed Twitter posts, intending to categorize them into extremist and non-extremist tweets. The framework comprises two popular deep learning algorithms: convolutional neural network (CNN) and LSTM. A CNN is used to extract relevant features and feed them as inputs into the LSTM model to capture the tweet’s sequential dependencies and global correlations by considering the previous data.

Mossie et al. [12] introduced an approach to detect hate speech against vulnerable minority groups on social media platforms. The Spark-distributed processing framework automatically collected social media posts and extracted features using word n-grams and word-embedding techniques such as Word2Vec to capture semantic relationships. Hate speech was identified using deep learning algorithms, mainly variations of RNNs such as gated recurrent units (GRUs). Word2Vec was utilized to group words associated with hate speech, allowing for the potential prediction of targeted ethnic groups. The experiment results demonstrated that using word-embedding techniques such as Word2Vec for feature extraction was highly effective for detecting hate speech for traditional machine learning and deep learning algorithms. The deep learning and traditional machine learning algorithms were fed with the extracted features from Word2Vec, and the GRU variant outperformed the other algorithms.

Asif et al. [13] proposed a new cyber-trolling detection method using a labeled troll and non-troll sample dataset. A graph convolutional network (GCN) is employed to model the relationships between words through cosine similarity thresholds, improving toxic content classification compared to traditional methods. The workflow includes data cleaning, TF-IDF feature extraction, graph construction, model training, and accuracy evaluation. However, the model’s effectiveness may be affected by the dataset’s domain coverage; new forms of trolling or platform-specific language may necessitate retraining to maintain accuracy.

Focusing on suicidal ideation detection in Arabic Twitter posts, Abdulsalam et al. [14] framed the task as a supervised classification problem. Data were collected via Tweepy using suicide-related keywords, manually annotated by two judges (including a cyberpsychology expert) to label tweets as ‘Suicidal’ or ‘Non-Suicidal’. Preprocessing (e.g., duplicate removal) and feature extraction steps preceded both the machine learning and deep learning models used for classification, culminating in performance analysis. The unique characteristics of Arabic dialects and selected suicide-related keywords may not capture all nuances of self-harm language. Additionally, the reliance on expert manual labeling presents challenges for scaling to larger datasets or diverse cultural contexts.

Daraghmi et al. [15] proposed a hybrid deep learning model for cyberbullying detection by combining CNN, Bi-LSTM, and GRU layers, implemented in Keras/TensorFlow with grid-searched hyperparameters. Stacked word embeddings (GloVe, FastText) captured semantic relationships, while CNN layers aided feature extraction, and Bi-LSTM plus GRU processed long-term dependencies efficiently. Dropout, max pooling, and batch normalization mitigated overfitting, and the Adam optimizer refined parameters over multiple epochs for binary classification. However, the computational complexity of maintaining numerous sequential layers required significant processing power and hindered real-time detection. Additionally, generalization to other languages or domain shifts required further data and retraining of embedding layers.

Our proposed model aims to integrate a variety of extensive metadata features. These metadata are combined with textual input processed through LSTM networks. Additionally, the ResNet–DNN sub-architecture refines the metadata representations before merging them with the textual embeddings. This approach may provide greater resilience to noisy or incomplete data compared to methods that rely solely on text. Our results demonstrate an improved accuracy of approximately 93.14%. Although this is only modestly higher than some baseline models, it can be critically important in forensic applications. Previous research shows that machine learning and deep learning have significant potential in social media forensics. However, most frameworks focus on either textual data or basic metadata, often ignoring the benefits of combining both. Our work fills this gap by introducing a method that integrates text and metadata in a deep learning pipeline.

3. Dataset Information

To conduct a thorough investigation, we require a detailed and all-inclusive dataset to serve as the basis of our analysis. Fortunately, Dadkhah et al. [16] recently released a real/fake social media content dataset. The dataset comprises over 180,000 labeled tweets, making it one of the most extensive datasets. The dataset was carefully selected using a three-factor active learning verification process. The tweets were labeled by 456 highly skilled Amazon Mechanical Turk workers. Additionally, the dataset includes three supplementary social media scores: influence score, Bot, and credibility. These aid in comprehending user characteristics and data patterns.

The dataset contains different features: the statement, author name, target truthfulness of the tweet, manual keyword, and labels indicating the tweet’s accuracy bias. Moreover, the dataset includes information regarding the text, vocabulary, metadata associated with each tweet, and information about users who posted it. The dataset specifically targets tweets concerning real or fake political news. It was crowd-sourced through Amazon Mechanical Turk to determine whether each tweet was real or fake news.

This dataset is ideal for our study as it contains many samples that can efficiently train the proposed hybrid deep learning model. Additionally, the labeling strategy used in the dataset is convincing and aligns with our study’s objective. Our proposed multi-input framework benefits from diverse features like textual and metadata information. This dataset aids the research in identifying markers that predict truthfulness by analyzing social media post patterns and nuances.

The dataset is available in three versions: one for binary classification and two for multi-class classification. The multi-class datasets contain three and five classes, respectively. We used a CSV file for binary classification to distinguish between fake and real tweets. There are sixty-four features in this version of the dataset, including the target class, the tweet, and the metadata information.

3.1. Dataset Preprocessing

To properly utilize our hybrid deep learning model, the data must be preprocessed. The following preprocessing phases were considered:

• Stop-word removal: Well-known words that do not add enough value in the context were ignored. These common words were removed to reduce noise and focus on more meaningful content. Additionally, special characters were eliminated from the tweets to standardize text format and reduce noise.
• Tokenization: The tweets were divided into separate units of words, known as tokens. This step converts unstructured text into discrete elements, enabling numerical representation and efficient processing by neural networks.
• Sequence padding: Input sequences were adjusted to a fixed length (50 tokens per tweet) by padding shorter sequences with zeros or truncating longer sequences to ensure consistency in data dimensions.

These preprocessing steps ensure the data are in a suitable format for training and testing the hybrid deep learning framework.

3.2. Dataset Division

The dataset is divided into training and testing sets to enable the training and validation of the model. The percentage of the data used as the training set is 80%, and the remaining 20% is designated for the testing set. This data split criterion ensures the unbiased validation of the model’s performance on unseen data.

4. Proposed Method

Figure 1 shows the methodology used to explore textual data and metadata as multiple inputs in a hybrid deep learning framework comprising LSTM, ResNet, and DNN layers. The LSTM submodel processes the tokenized text using standard gating Equations (2)–(6). At the same time, the metadata are processed by ResNet and DNN blocks, which utilize skip connections to reduce vanishing gradients and generate robust feature embeddings. The outputs from these two streams are then concatenated and fed into a softmax classifier, producing the final classification of suspicious or normal. The text undergoes preprocessing, and metadata are enriched with additional features. The metadata features are input into the random forest feature selection algorithm to choose the most significant features. LSTM is used for sequential text, while a combination of ResNet and DNN layers handle the metadata. The goal is to utilize a multi-input model with features of varying characteristics and shapes.

4.1. Feature Selection Methods

The deep learning framework’s performance can be improved by refining the metadata’s 64 features for accuracy and detection time. The metadata are passed through the feature selection mechanisms of the random forest ensemble method and feature correlation technique. As shown in Algorithm 1, the random forest training algorithm uses bootstrap aggregating (bagging) to train the tree learners. Having a training set, $X=x_1,\cdots ,x_i$, with labels $Y=y_1,\cdots ,y_i,$ in a recursive manner, bagging chooses a random sample with replacement N times from the training set and trains trees to learn these samples:

Algorithm 1 The random forest training algorithm used to refine the 64 metadata features

for $n=1,\cdots ,N$: do   1. Sample with replacement i training samples from $X,Y$; name these samples $X_n,Y_n$   2. Fit a random forest tree $f_n$ on $X_n,Y_n$. end

After completing the training process, we can predict the outcome of unseen samples $\acute{x}$ using the plurality vote of the random forest trees. The trained random forest ensemble method consists of twenty trees. The maximum depth is the number of splits each tree can make; we use ten as a maximum depth to avoid underfitting (if we choose a small number of less than five) and overfitting (if we choose a large number of more than ten). After selecting the thirty-five most significant features, these metadata characteristics will be fed into the proposed hybrid deep learning system.

We also use correlation heatmaps to visualize the strength of the relationship between numeric variables. These heatmaps are utilized to learn which features are related to each other and the strength of the relationship between these features.

The metadata information used for the DNN–ResNet submodel includes the features selected by the random forest feature selection and correlation mechanism and the features chosen by either of these methods.

4.2. Pre-Trained Word Embedding

Pre-trained models are becoming increasingly popular for various classification tasks because they reduce the time and effort required to train a deep learning model from scratch [17,18]. The pre-trained models have been trained on vast amounts of data, allowing them to recognize patterns and relationships within the data, making these models the perfect fit for NLP tasks such as text classification [19].

Global Vector (GloVe) [20] is a pre-trained word-embedding model trained on a large corpus representing words as vectors. These vectors serve as contextual words to enhance text classification performance and generalizability [21]. We used a pre-trained GloVe word-embedding model that generated a 50-dimensional word-embedding vector. Each n-word, represented as $W(w_1,\cdots ,w_n$), is input and converted into a vector of a specific dimensions, d. The dimension space of each word is $S^d$; therefore, each input text is represented as $S^{n\times d}$. Thus, the feature vector can be represented as $c_v$ and the word is concatenated in the following manner:

$$c_v=w_1\oplus w_2\cdots \oplus w_n$$

(1)

This pre-trained word-embedding model can handle the dimensional problem of high-dimensional sparse matrices.

4.3. LSTM for Tweet Handling

An LSTM network comprises interconnected neural network units. Every unit keeps track of its previous state information. This information is stored, written, or read from a cell, which functions like memory. The cell controls the storage, reading, writing, and removal of information by opening and closing gates. The cells act by filtering data based on their strength and importance through a set of weights once they receive signals. The LSTM architecture typically comprises input, hidden, and output layers [22]. The input gate is represented by $i_g$, the output gate is represented by $o_g$, the forget gate is represented by $f_g$, and the cell state is represented by ${\tilde{L}}_g$. The gate responsible for deciding which information to keep and discard from the cell state, ${\tilde{L}}_g$, is the forget gate, $f_g$. The logistic function outputs a value between 0 and 1 to make a decision. The logistic function can be denoted as

$$f_g=\sigma (W_f[o_{t-1},i_t]+B_f)$$

(2)

where $\sigma$ represents the activation function, $W_f$ represents the forget gate weight, $B_f$ represents the forget gate bias, $i_t$ represents the input at time t, and $o_{t-1}$ represents the hidden layer output at time $t-1$.

The LSTM input gate $i_g$ updates input values based on its decision with the help of the LSTM blocks. The calculation at the input gate $i_g$ is performed as follows:

$$i_g=\sigma (W_i[o_{t-1},i_t]+B_i)$$

(3)

where the bias of the input gate is represented by $B_i$. Additionally, the calculation at the cell state ${\tilde{L}}_g$ is performed as follows:

$${\tilde{L}}_g=\tanh (W_l[h_{t-1},i_g]+B_l)$$

(4)

The cell weight is denoted by $W_l$, and cell bias is represented by $B_l$.

The output is the final state; as can be seen in the following equation, the output gate $o_g$ contains a tanh activation function to help determine which part of the cell state is the outcome.

$$o_g=\sigma ({\mathrm{W}}_{\mathrm{o}}[{\mathrm{h}}_{\mathrm{t}-1},{\mathrm{i}}_{\mathrm{t}}]+{\mathrm{B}}_{\mathrm{o}})$$

(5)

The output gate weight is represented by $W_o$, and the output gate bias is denoted by $B_o$. The hidden gate $h_g$ calculation is shown in the following equation. The output gate $o_g$ is multiplied by a logistic function to produce a value between 0 and 1.

$$h_g=o_{\mathrm{g}}+\tanh \left({\tilde{L}}_g\right)$$

(6)

Our model uses an embedding layer that maps a sequence of word indices to vectors; this layer receives fifty vocabulary sizes and fifty embedding matrices, representing the weights. The embedding layer is connected to an LSTM layer that produces an output of 128 units. Then, a dense layer receives the output of the LSTM layer and uses relu as an activation function to produce an output of ten units. This is the first input to be fed into our multi-input and hybrid model.

4.4. ResNet and DNN for Metadata Handling

The combination of the ResNet and DNN layers handles the metadata that have been selected using the two feature selection methods (random forest and correlation feature selection methods). The ResNet consists of building blocks employed to solve the gradient fading issues [23]. To attain high performance, it is necessary to create a deep network. The ResNet’s depth can reach more than 1000 layers. This study aims to achieve high accuracy while also considering efficient detection time. Therefore, our ResNet is not very deep. The building block of two stacked ResNet layers can be defined as follows:

$$O\left(i\right)=F(i,\left\{W_r\right\})+i,$$

(7)

where i represents the input vector of the building block and $O\left(i\right)$ denotes the output vectors of the building block. The residual mapping to be learned is denoted as $F(i,\left\{W_r\right\})$.

The DNN relies on statistical models inspired by biology [24]. It comprises interconnected neurons and nodes that use the multilayer method for perception. The DNN is a non-linear model that takes input and produces output values. The structured nodes of neurons are connected systematically to construct randomly interconnected layers. The typical DNN consists of three layers: input, hidden, and output. Nodes are assigned numeric weights through input and output processes, which are then transformed by an activation function.

$$D={\displaystyle \frac{1}{2}}\sum _{n=1}^N{\left(o_i-r_i^I\right)}^2$$

(8)

where $o_i$ represents the output and $r_i^I$ denotes the current responses at the node i of the output layer. On the other hand, N refers to the number of nodes the output layer consists of. As shown in the following equation, this method is used iteratively to correct parameter weights by computing and adding them to previous outputs.

$$\left\{\begin{array}{c}{\displaystyle \mathrm{\Delta }{w}_{a,b}=-\mu \frac{\partial E}{\partial w_{a,b}}}\hfill \\ {\displaystyle \mathrm{\Delta }{w}_{a,b}\left(e+1\right)=\mathrm{\Delta }{w}_{a,b}+\alpha \mathrm{\Delta }{w}_{a,b}\left(e\right)}\hfill \end{array}\right.$$

(9)

where $w_{a,b}$ refers to the weight parameters of nodes a and b and $\mathrm{\Delta }$ represents the positive constant known as the learning rate, used to regulate the adjustment to be performed. The momentum factor is denoted by ∂ and can be 0 or 1. The iteration number is referred to by e, and the smoothing factor $\alpha$ is used to smooth the changes between weights.

Our DNN–ResNet submodel starts with a dense layer that receives the metadata input of 42 features and produces an output of 64 units. Then, a relu activation function is applied to apply the non-linearity to the output. Subsequently, the output is passed through a batch normalization utility to enable the natural network to train more quickly and more stably. Then, a ResNet block is applied, which consists of two dense layers, two relu activation functions, dropout layers with a rate of 0.5, and two batch normalization methods. The output of the ResNet is 32 units, representing the second input of our multi-input and hybrid framework.

4.5. Multi-Input Concatenation

The last step of our proposed method is to combine the textual information with the metadata in our hybrid model. In this step, the output of the LSTM submodel (i.e., the ten units) is concatenated with the output of the DNN–ResNet submodel (i.e., the 32 units). These outputs become the inputs to our multi-input framework, which produces an output of 42 units after passing the inputs into the concatenation layer. Then, the output is passed through a dense layer that produces an output of 16 units. Finally, the softmax activation function is applied to classify the samples into two classes (i.e., suspicious and normal).

5. Results and Discussion

Figure 2 displays how removing stop words positively affects the correlation of a sequence of meaningful tweets. The top five biagrams before removing the stop words are shown in Figure 2a; the inference of meaningful sentences is hindered. After removing the stop words, the correlation between word sequences helps NLP models learn context, as shown in Figure 2.

Both feature selection methods selected forty-two features. Thirty-five features were selected using the random forest selection method, while twenty-nine features were selected using the correlation method. These features have positive or negative correlations, as shown in Figure 3.

As can be seen in the figure, the correlation heatmap contains twenty-nine features; each feature is represented as a column, and rows show the relationship between each pair of features. Each cell (the intersection of the feature and its counterpart) includes a value (indicated by blue and red colors in this plot) that indicates the strength of the relationship. Positive values denote a useful relationship between the two features, while negative values represent an imperfect relationship between the two features.

Hyperparameter optimization is critical in enhancing deep learning architectures, especially for multimodal setups, where textual (LSTM-based) and numerical features must be modeled together. We employ Keras Tuner’s Bayesian optimization approach to systematically search over an extensive range of hyperparameters, including LSTM configuration (i.e., number of units, bidirectionality, number of layers), trainable or static embeddings, dropout levels for the numeric branch, and various optimizer settings (i.e., Adam and RMSprop) with different learning rates. Early stopping is used to avoid overfitting and to identify near-optimal parameters without exhaustive training.

Table 1. Comparison between the baseline proposed method and the hyperparameter-optimized model.

Aspect	Baseline Proposed Method	Hyperparameter-Optimized Model
Overall Framework	This is a multi-input, hybrid deep learning system (merging LSTM, ResNet, and DNN). It focuses on text, utilizing LSTM and metadata, and employs ResNet and DNN to identify fake tweets.	It has an identical multi-input structure, but key hyperparameters (e.g., LSTM size, dropout, optimizer) are systematically tuned for higher performance.
Feature Selection	It utilizes random forest and correlation-based feature selection to reduce 64 metadata features to the top 35, enhancing computational efficiency.	It employs an identical feature selection approach to the baseline, but the final architecture is optimized after feature filtering.
LSTM Configuration	It uses a single LSTM layer with 128 units, followed by a dense layer with 10 units.	It dynamically tunes LSTM size (i.e., 64–256 units), number of LSTM layers (i.e., 1 and 2 layers), and optional bidirectional setting.
Metadata Modeling	It employs a ResNet block with two dense layers, skip connections, and DNN layers (up to 64 units). It also employs batch normalization and dropout.	It retains ResNet and DNN approaches but further optimizes layer sizes, dropout rate (0.0–0.5), and activation functions through Bayesian search.
Embedding Layer	It employs pre-trained 50-dimensional GloVe embeddings (non-trainable).	It toggles between trainable and non-trainable embeddings, letting the tuner decide the best approach.
Optimizer and Learning Rate	It uses a fixed choice (e.g., Adam, LR = $1\times {10}^{-3}$ ) without a large-scale hyperparameter search.	It explores Adam and RMSprop with multiple learning rates ($1\times {10}^{-2}$, $1\times {10}^{-3}$, $1\times {10}^{-4}$), picking the best via validation.
Training and Early Stopping	It uses standard training loops with a preset number of epochs and early stopping heuristics, but it does not perform a systematic search.	It employs a tuner (Bayesian optimization) with early stopping and searching of a wide hyperparameter space.

highlights the main distinctions between the baseline proposed method [7] and the hyperparameter-optimized proposed method. The former describes a multi-input, hybrid deep learning framework (i.e., LSTM, ResNet, and DNN) with random forest and correlation for feature selection; the latter emphasizes a Bayesian optimization process that systematically tunes key model parameters (e.g., LSTM units, dropout rate, optimizer type). The proposed method focuses on architectural design for jointly handling text and metadata, whereas the hyperparameter-optimized proposed method concentrates on searching for the best hyperparameters to bolster the overall performance.

This search identified a configuration that yielded a significant performance improvement, with faster convergence and higher overall accuracy than our initial baseline. Specifically, the final model converged on a single LSTM layer with 64 units, a trainable GloVe embedding, 0.3 dropout on the numeric branch, and an Adam optimizer with a 0.01 learning rate. Figure 4 illustrates the training and validation performance for this best set of hyperparameters, plotting loss on the left y-axis with blue curves and accuracy on the right y-axis with red curves. Solid lines represent the training metrics, while dashed lines represent the validation metrics.

Notably, the model’s loss curves rapidly descend within the first few epochs, while the accuracy curves, especially for validation, rise accordingly. Training and validation losses are below 0.5 by the third epoch, and validation accuracy stabilizes near 0.9 in subsequent epochs. Observing that the training and validation curves remain relatively close, this indicates that the model does not suffer from severe overfitting under these hyperparameter settings. The training and validation curves are closely aligned, but validation accuracy begins to diverge slightly after the tenth epoch, indicating a smal amount of overfitting. This suggests that the model may require periodic retraining or updates with new data to maintain its performance in real-world applications. After final tuning, testing on an unseen set yields a test loss of 0.2299 and a test accuracy of 93.14%, demonstrating that systematic hyperparameter optimization can substantially boost performance in this multimodal classification task.

As shown in

Table 2. Comparison of the proposed model and the rest of the algorithms.

Algorithm	Accuracy
LSTM using textual data	89.17%
BERT using textual data	83.80%
RoBERTa using textual data	84.40%
DNN for metadata (three features) and LSTM for text	79.54%
Baseline proposed model	90.62%
Hyperparameter-optimized proposed model	93.14%

, the proposed multi-input and hybrid model effectively identifies fake or real tweets using two data sources. Using the LSTM for textual data enables the framework to learn the text semantics; the LSTM achieves good performance (i.e., an accuracy of about 89%). However, the performance could be improved when utilizing the metadata and textual data. As shown in the table, the experimental results show slight improvement over LSTM, yielding an accuracy of 93.14% (i.e., about four percent) when combining text and metadata, which might add value in identifying patterns and relationships that can be used in forensics investigations. To ensure our model effectively distinguishe fake posts from real ones, particularly under class imbalance, we evaluated our model using precision, recall, F1-score, and AUC-ROC. Our results show a macro precision of 0.93, macro recall of 0.93, macro F1 of 0.93, and AUC-ROC of 0.95. This performance indicates that the classifier consistently identifies both real and fake samples.

To assess the computational resources required, we compared both the training and inference times of the baseline and hyperparameter-optimized models. All experiments were executed using a Google Colab environment (CPU with high-RAM runtime). Both models were trained for up to 10 epochs with early stopping enabled. The baseline model took approximately 227.65 s to complete training and 12.57 s for inference (testing on the held-out test set). The hyperparameter-optimized model trained in 104.55 s and completed inference in 7.22 s. This cut the training time by 54% and inference time by 43%. These improvements show that optimizing the hyperparameters boosts predictive accuracy and makes the model more efficient. This is essential when deploying the model in social media forensic applications.

We also compared our model with two transformer-based models, specifically BERT and RoBERTa. These two models have considerable computational overhead compared to our proposed hybrid framework. Training BERT using the same settings required 7215.88 s (about 2 h), while RoBERTa training took slightly longer, approximately 7530.69 s (about 2.1 h). The inference durations for BERT and RoBERTa were also relatively high, being 49.83 s and 50.99 s, respectively. While the transformer-based models achieved competitive accuracy (BERT: 83.80%, RoBERTa: 84.40%), our hybrid model attained a superior accuracy of 93.14% (and the LSTM model utilized to classify text achieved 89.17%), highlighting that it is not only efficient in terms of computational resources but also effective in the task of authenticity verification on social media posts.

While ResNet was initially designed for image-based feature extraction, the concept of skip connections to stabilize gradient flow can be beneficial beyond computer vision tasks. In our system, the metadata submodel does not employ a deep, full-scale ResNet architecture; we utilize lightweight residual blocks comprising dense layers with skip connections. This approach helps mitigate vanishing gradients and has proven effective for learning from tabular data without introducing excessive model depth or complexity. We implemented a TabTransformer on our numeric metadata to compare it to our metadata submodel. The empirical results show that the ResNet–DNN approach outperformed the simpler TabTransformer configuration. Specifically, ResNet–DNN achieved a test accuracy of 75.45%, whereas TabTransformer reached only 51.34% accuracy. This indicates that the complexity introduced by multi-head self-attention in TabTransformer did not automatically yield better performance under our current setup and dataset. We acknowledge that TabTransformer and other specialized architectures can effectively handle categorical embeddings and attention-based feature tokenization. However, they often involve additional overhead and tuning factors that may not pay off on simpler or smaller tasks without more extensive optimization. We plan to investigate domain-specific modifications (e.g., advanced embeddings, more refined hyperparameter tuning, or specialized data preprocessing) to see whether the added complexity of attention-based methods can produce substantially improved performance. From these initial attempts, the lightweight ResNet–DNN design provides a strong baseline with less computational burden, supporting the idea that skip connections can be beneficial even for structured numeric data.

Limitations and Future Work

Our hybrid framework is promising. However, some limitations need to be addressed to improve its practicality. Limitations concerning practical implementation and dependence on expert knowledge should be resolved. The model’s effectiveness relies heavily on labeled datasets from domain experts, raising questions about the criteria for classifying posts as real or fake, potential subjective judgments, and variability in annotator expertise. The model’s accuracy depends on the quality and timing of its data. We need to update and retrain the model regularly to keep up with changes in social media language and misinformation trends. We also need clear guidelines that involve cybersecurity experts and linguists to ensure reliable authenticity labels. Adding new annotated data every three months will maintain the model’s reliability, and utilizing online methods for continuous learning can help update the model dynamically as novel misleading content emerges.

The evaluation results show that there are minimal signs of overfitting, but the model may still be too specialized on the current dataset. This could make it hard for the model to work well with real-world data or adjust to information from platforms like Instagram, Reddit, or Facebook, where text and metadata are different. Future work would focus on periodic retraining, transfer learning, updating expert-curated labels, and using diverse datasets across platforms to ensure the model remains effective as language and platform characteristics evolve. Future work will assess our model’s robustness and generalization using larger, more diverse datasets with comprehensive text and relevant metadata. While testing the model on additional real-world, out-of-dataset examples would provide valuable insight into practical performance, obtaining datasets with suitable and representative metadata is currently challenging due to availability and privacy constraints. Nevertheless, future studies will prioritize acquiring and annotating external real-world samples to assess the model’s ability to generalize beyond the original dataset context.

There are also some concerns regarding privacy and ethics when it comes to collecting metadata under GDPR. We need to ensure that data are anonymized, obtain user consent, and be transparent about how we use data. The black-box nature of deep learning makes it hard to explain how our models make predictions, which can lower user trust. Future work should include explainable artificial intelligence (XAI) techniques to improve transparency and help users understand the models better. Additionally, concerns regarding biases in the dataset might exist. Models trained on a single dataset concentrate on certain traits that are not deployable through other social media platforms, which can cause degraded prediction performance. Also, labeling biases, time restrictions, and differing expert beliefs can lead to reduced reliability.

Moreover, the utilized dataset includes textual and numeric metadata but lacks visual modalities required by advanced multimodal transformer architectures such as ViLT and MMBT. These models integrate visual and textual information and have shown potential in social media analysis tasks involving images and text. Future work would consider collecting a dataset containing images along with text and metadata, thus enabling an evaluation of multimodal transformer models to determine if incorporating visual modalities further enhances performance in authenticity detection tasks.

6. Conclusions

In this study, combining text and metadata was found to be effective for data forensics, especially for social media platforms like Twitter. The hybrid deep learning model combines LSTM for text analysis with ResNet and DNN for metadata, achieving an accuracy of 93.14% in identifying fake content. This demonstrates the importance of multi-input pipelines in managing diverse social media data. The preliminary results are promising, showing high accuracy in identifying fake content. However, this study needs to be improved to optimize/enhance the proposed framework. We will investigate another dataset for better generalization and try new features to improve performance. Additionally, we will develop a tool that monitors social media posts and detects fake content in real time.