A Bio-Inspired Grey Wolf Approach to Enhancing Fake Profile Detection in Online Social Media

This section reviews various approaches for identifying fake profiles on OSNs, highlighting the use of numerous ML algorithms.

The framework for detecting fake profiles on OSNs [5] employs open-source big data tools and Long Short-Term Memory (LSTM) networks for analysis. It integrates the Dispersive Flies Optimization (DFO) metaheuristic to enhance feature selection from the dataset. Additionally, the approach underscores the importance of ethical data collection, taking into account both public and private user attributes.

The issue of detecting fake profiles on Facebook was addressed through a hybrid methodology [6]. This approach involves a two-phase process: first, the Satin Bowerbird Optimization Algorithm (SBO) is used to establish initial clusters and determine optimal centroids for profile classification. Then, the K-means clustering algorithm is applied to categorize each profile as either real or fake.

A bio-inspired algorithm called the Fire Hawk Optimizer (FHO) was introduced to address the issue of fake profile detection [7]. Various feature groups from a Twitter dataset were evaluated to assess their effectiveness in identifying fake profiles. Based on their findings, they recommended the use of the Gradient Boosting Classifier (GBC) as the Fitness function.

The research [8] dealt with the problem of fake profiles on social media by introducing a novel method that integrates various ML algorithms to evaluate user behavior and profile information. This approach, termed "ensemble," employs a Majority Voting Technique (MVT) to classify profiles as either fake or real. The findings indicate that this method holds significant potential for improving the security of social media platforms.

Machine learning algorithms were applied to detect fake profiles on OSNs, with multiple models evaluated for effectiveness in identifying fraudulent accounts [9]. The study emphasized the importance of diverse evaluation approaches, such as confusion matrices and error rate analyses, to determine the optimal model.

In the context of existing research, Mahammed et al. [10] assessed how effectively the Fire Hawk Optimizer (FHO) detects fake social media profiles. The authors tested FHO's performance. Their goal was to identify the most effective feature subsets from a Facebook dataset to differentiate between real and fake profiles.

A novel method was introduced utilizing an extensive feature set to examine profile information, network connections, and user behavior for fake profile detection. An adjustable Bagged Tree Algorithm (BTA) was also proposed, enhancing decision tree models by eliminating irrelevant branches to improve accuracy and efficiency [11].

An ML model combining Logistic Regression (LR) with Gradient Descent Optimization algorithm (GBO) was introduced to identify fake profiles on OSNs [12]. When tested on an Instagram dataset, the model demonstrated strong performance.

Raghavendra et al. [13] presented detailed solution framework for detecting and eliminating fake profiles on OSNs. This solution utilizes ML algorithms to examine different aspects of user behavior and identify related anomalies. It highlights the importance of user education, ongoing updates, and privacy-respecting practices. The approach incorporates a variety of algorithms, including ANN, RF, Extreme Gradient Boost (XGBoost), LSTM, and a Voting Classifier (VC).

Table 1 offers a detailed synopsis of the latest advances in detecting fake profiles across different social media using ML algorithms.

Table 1. Related work synopsis

Table 1 provides a comprehensive overview of the information. ML: Machine Learning, Meta: Metaheuristic, Spec: Specification, Combo: Combination, Compo: Composition, Acc: Accuracy. It indicates that various OSNs are employed in the detection of fake profiles, with Facebook being the most frequently examined platform, followed by Twitter and Instagram. GBC and LSTM are the most commonly used ML algorithms, each has demonstrated strong performance in detecting fake profiles. Other algorithms employed less frequently but still contribute valuable insights.

Metaheuristic techniques, particularly bio-inspired algorithms, are less common compared to traditional ML algorithms. FHO, SBO and BTA are featured in numerous studies. Although less prevalent, these metaheuristics offer alternative approaches to improving detection accuracy.

Hybrid and combination techniques are prominent, with three studies using hybrid methods that integrate various ML algorithms and bio-inspired algorithms. These approaches often lead to high accuracy rates, reflecting their effectiveness. Additionally, two studies employ combination techniques to enhance detection accuracy.

Dataset sizes in these studies vary, from relatively small datasets of around a thousand to 7,500 accounts to much larger datasets, such as 17,350 accounts [7]. Despite this variability, the results are notable for their high accuracy.

Table 1 clearly demonstrates that both ML and bio-inspired algorithms play decisive roles in the detection of fake profiles on OSNs. ML algorithms such as GBC is highly effective and widely used, achieving high accuracy rates in distinguishing between fake and real profiles. Bio-inspired algorithms, though less common, offer valuable optimization capabilities and can significantly enhance detection systems when combined with traditional ML algorithms.

The present work is positioned at the intersection of ML and bio-inspired algorithm combinations for fake profile detection, as deduced from Table 1. The materials and methods section introduces the proposed model based on the well-known bio-inspired GWO, utilizing an ML algorithm as the Fitness function for feature selection. The model is then applied to different datasets, using various ML algorithms to assess its performance.

This section presents the experimental results and offers a detailed discussion on the performance of the proposed bio-inspired algorithm in detection fake profiles. The results are evaluated based on metrics such as Accuracy, Precision and convergence, and efficiency also. Additionally, comparison and analysis of key parameters affecting the exploration/exploitation balance of GWO is presented.

Table 9 presents the confusion matrix for the different datasets used in the experiment. It provides valuable insight into the effectiveness of the proposed model which coupled with GWO and its LR-based Fitness function.

•The confusion matrix presents a good classification performance of both real and fake profiles:

•1025 true positives and 200 true negatives with Facebook dataset.

•464 true positives and 498 true negatives with Twitter dataset.

•326 true positives and 328 true negatives with Instagram dataset.

Table 9. Confusion matrix with GWO

This implies that only 18, 35, and 20 profiles are detected as false negatives for Facebook, Twitter, and Instagram, respectively. It demonstrates balanced fake profiles detection. These results highlight the model's Precision at 0.97, 0.93, and 0.94 for Facebook, Twitter, and Instagram, respectively (Table 10), meaning it is highly effective at minimizing false positives, with the best results achieved using the Facebook dataset. With Recall values of 0.96, 0.93, and 0.94 (for Facebook, Twitter, and Instagram, respectively), the model also performs well in detecting a significant proportion of fake profiles.

As shown in Table 9, GWO played a key role in enhancing the model’s performance. By mimicking the hierarchical hunting strategies of grey wolves, which is essential in optimization tasks like feature selection for fake profiles detection. Thereafter, through the iterative process of position updates guided by alpha, beta, and omega wolves, GWO effectively identifies the most relevant features. This ensures that the model focuses on the most distinguishing characteristics of real and fake profiles, contributing to its high classification Accuracy of 0.97, 0.93, and 0.94 for Facebook, Twitter, and Instagram, respectively (Table 10). Logistic Regression (LR) serves as the fitness function within GWO algorithm to assess how well a set of selected features can classify profiles correctly. This is done by evaluating the quality of feature subsets during the optimization process. LR is particularly well-suited for binary classification tasks, such as distinguishing between real and fake profiles. It is also important to point out that LR's natural ability to handle regularization helps control overfitting, as demonstrated across various datasets with different features and sizes.

Table 10. Results overview

Note: Dataset https://www.kaggle.com/free4ever1/instagram-fake-spammer-genuineaccounts

Table 10 compares the results obtained by the proposed model with those from studies providing datasets from OSNs such as Facebook, Twitter, and Instagram. The performance metrics evaluated include Accuracy, Precision, Recall, and F1-score. The results can be interpreted in terms of GWO's performance, including its use of LR as a fitness function, its exploration and exploitation capabilities, and its convergence behavior across the diverse solution spaces presented by the datasets.

For the Facebook dataset, the combination of GWO and LR emerges as a highly effective solution for fake profile detection. GWO's capabilities, coupled with its rapid convergence, enable it to select the most relevant features for LR, resulting in superior performance across all metrics. In spite of that, the previously reported model [14] achieved a slightly higher scores with DT with Accuracy, Precision and Recall reaching 0.97, 0.98 and 0.98 respectively compared to 0.97, 0.96, 0.96 obtained by the proposed model. This suggests that while DT might capture more true positives, it also introduces more false positives, which GWO mitigates. However, the proposed model surpasses the previous model's performance in terms of F1-score, achieving 0.95 compared to the earlier result of 0.93. It is worth to notice that the previous reported model [14], which employed SVM and K-NN showed lower Accuracy, Precision, Recall, and F1-score compared to GWO, indicating that they are less effective for this dataset.

On Facebook dataset, the high Precision (0.97) and Recall (0.96) demonstrate that GWO effectively balances the search for new solutions with refining existing ones. Its exploration ability allows it to search through a large solution space, identifying relevant features that may be scattered across various behavioral patterns, interaction data, and content attributes. Once promising feature subsets are found, exploitation ensures that the algorithm reduces its scope of action on refining the most optimal feature set.

On Facebook dataset, GWO exhibits rapid convergence, as showed by its Accuracy score of 0.97. This rapid convergence enables the proposed model to efficiently identify the most relevant feature subsets, leading to superior Precision and Recall, which is valuable when working with a dataset like Facebook's.

For the Twitter dataset, GWO/LR combination proves to be a very efficient as solution for the detection of fake profiles. GWO's capabilities (exploration, exploitation and rapid convergence), facilitate its identification of the most relevant features for LR, consistently outperforming in all aspects of evaluation. While the previously reported model [15] achieved an equal Recall score of 0.93 obtained by NB, the proposed model consistently outperforms it in terms of Accuracy (0.93 vs. 0.91) and Precision (0.93 vs. 0.88).

On Twitter dataset, in order to avoid overfitting, GWO\LR's capabilities ensure that selected features generalize well. GWO’s exploration is particularly important, given the rapid pace and diversity of user activity. The model excels at analyzing dynamic data (interactions, mentions and retweets) to find important features. Then, GWO’s exploitation is used to refine the feature set, as shown by the Precision and Recall scores of 0.93.

On Twitter dataset, GWO’s convergence is also evident in its Accuracy score of 0.93, which surpasses other algorithms. The proposed model converges toward a solution that balances false positives and false negatives effectively, achieving the highest F1-score of 0.93. It is clear that the proposed model has the ability to handle dynamic and varied data as Twitter dataset.

1a.png

(a)

1b.png

(b)

1c.png

(c)

Figure 1. GWO’s convergence performance, (a) accuracy, (b) precision, (c) recall

Figure 1 visually illustrates GWO’s convergence behavior over iterations in terms of accuracy, precision, and recall, clearly outperforming other ML algorithms. The convergence results, illustrated in Figure 1, highlight the effectiveness of the GWO-based approach compared to traditional machine learning (ML) algorithms for fake profile detection on social media. In Figure 1(a), the accuracy of GWO rises steadily from 80% to 97%, clearly outperforming baseline ML models that plateau at 85%. This improvement reflects GWO’s capability to guide effective feature selection, resulting in more accurate classification. Similarly, Figure 1(b) shows that GWO achieves precision values exceeding 96%, compared to ~82% for standard ML. This indicates a marked reduction in false positives, enhancing the model’s reliability in distinguishing fake profiles from real ones. Figure 1(c) shows recall improving from 77% to over 95% with GWO, highlighting its strength in reducing false negatives. This is crucial in security contexts where missing fake profiles can lead to harmful consequences.

A detailed error analysis across the three datasets—Facebook, Twitter, and Instagram—revealed distinct patterns in misclassification. The majority of false negatives occurred within the Instagram dataset, largely due to sparse and inconsistent profile attributes, which limit the model’s ability to extract distinguishing features. Unlike Facebook and Twitter, where users tend to share more structured and informative data (e.g., friend counts, bio text, post activity), Instagram profiles often lack depth, making fake profiles harder to differentiate from real ones. In the Twitter dataset, most false positives were associated with verified or bot-like real accounts that exhibited abnormal behavior, such as excessive posting or retweeting, mimicking fake profiles. For Facebook, errors were more balanced between false positives and false negatives, often due to duplicated or minimalistic user information, which blurred the decision boundary. These findings suggest the need for platform-specific feature engineering and the possible integration of behavioral or temporal features to further improve model accuracy, especially in data-sparse environments like Instagram.

Also, Table 10's results can be interpreted in terms of datasets specifications. Each dataset provides distinct sets of features that highlight the behavioral patterns and profile characteristics of users. These features, along with the distribution of real and fake profiles, directly influence the model’s ability to detect fake profiles.

Size and distribution: Facebook dataset has a much larger proportion of real profiles compared to fake ones (16%), which could make detecting fakes more challenging. When, Twitter and Instagram datasets are perfectly balanced, which allows for better performance evaluation in fake profile detection as the algorithm has equal exposure to real and fake profiles.

Features set: Facebook dataset focuses on profile information and social activity, which gives a lot of data related to social behaviors but lacks some of the detailed engagement metrics (hashtags and mentions). Whereas, Twitter highlights account behavior, including follower-following ratios and tweet activity, which provides rich behavioral data that could help for detecting fake profiles. While, Instagram focuses more on profile creation characteristics, such as username, bio length, and followers-to-follows ratio, as well as private settings and presence of an external URL. This could make it particularly suitable for detecting fake profiles, such as those created automatically.

Data imbalance: Facebook dataset presents an imbalanced dataset where the real profiles significantly outnumber the fake ones (5:1 ratio), which can introduce biases in training models. Twitter and Instagram datasets, being balanced datasets, do not suffer from this issue, allowing algorithms to train on an equal representation of real and fake profiles.

Complexity features: Facebook's feature set is broader in terms of social behavior (groups, friends, education, check-ins), making it a more complex dataset for models that need to understand diverse aspects of user interaction and profile details. In turn, Twitter dataset adds complexity with its focus on engagement metrics, such as hashtags, mentions, and tweet activity (behavior-based features), which are essential to understanding how users engage with the platform. Yet, Instagram dataset is a mix of both profile structure and activity. Its emphasis on profile features (e.g., numerical characters in usernames) could make it easier to detect fake accounts, but also lacks some of Twitter's engagement data.

Despite the promising results achieved by the proposed model, certain shortcomings persist. Therefore, it is imperative to propose solutions to address these issues.

Additional evaluation metrics: While Table 10 focuses on common metrics like Accuracy, Precision, Recall, and F1-score, other evaluation metrics [26], such as Area Under the ROC Curve (AUC), Matthews Correlation Coefficient (MCC), or Specificity, could provide deeper insight into model performance, especially in imbalanced datasets.

Incorporating more advanced feature engineering: The datasets (especially Facebook) include various profile-related features, but additional behavioral features and text analysis (e.g., sentiment analysis of posts or descriptions [27]) could help improve the algorithms’ ability to differentiate between real and fake profiles.

Balancing false positives and negatives: While some algorithms achieve high Recall, they suffer from false positives (low Precision). Cost-sensitive learning techniques [28] could help minimize false positives by adjusting the classification threshold or weighting false positives more heavily in the loss function.

Addressing imbalance between precision and recall: Several models, especially NB on Instagram dataset, show high Recall but suffer from lower Precision. This indicates that while true positives are identified well, false positives remain problematic. A solution to overcome this situation is to go for hybrid models.

Algorithm-specific tuning: Certain algorithms, like K-means, consistently underperform and may require specialized tuning or adjustments for binary classification tasks. Applying feature selection techniques or exploring enhanced clustering methods (e.g., SK-NSGAII [29]) could help improve K-means’ performance. The exploration of additional metaheuristics, not only bio-inspired optimization algorithms [30], but also those inspired by different approaches [31], aims to enhance performance.

It is important to acknowledge that the iterative feature selection process in GWO increases computational runtime compared to traditional ML methods. However, this additional computational overhead is justified by the substantial improvements in classification accuracy.