Based on data that can be found on social networks, the information is separated into four categories: hyperlinks, images, audio, and text (a subset of spoken language primarily produced with a text or string to examine the content) [11]. OSNs are receiving attention from users who are malicious or abnormal and engage in malicious activities such as harassing others, plotting attacks (in which terrorists may be involved), and disseminating false information [12]. Spam is the term for unsolicited messages that are sent in large quantities by fostering a sense of community trust. Spammers engage in illegal acts including phishing, advertising, surveillance, assault against women, and cyberbullying, among others [13]. Instead of using legitimate accounts, spammers typically distribute spam using fraudulent, compromised, or cloned accounts, crowd-sourcing strategies, and automated bots [14]. The taxonomy of various social spam detection techniques and approaches are observed as follows: URL list-based spam filtering (Blacklist, Whitelist, Greylist), honeypot and honeynet-based spam detection, and machine learning (ML) and deep learning (DL)-based social spam detection. ML and DL are used for social spam content detection including malicious URL detection [15] and text-based spam detection [16,17].

Social media bots (SMBs) are tools that people and organizations employ to spread information, expand their reach, and boost their impact. Malicious bots can annoy or burden users by participating in unethical actions, including stealing the identities of real users, persuading voters to favor politicians, spreading hate speech, and other divisive material [18]. SMBs are classified into three main groups: benign bots, neutral bots, and malicious bots. For SMB detection, the most used ML methods are random forest (RF), SVM, and AdaBoost, while LSTM and CNN are the two most widely used DL algorithms; unfortunately, there is a lack of large datasets to train models [18].

In [19], bidirectional encoder representations from transformers (BERT) are proposed. It has been shown that the pre-training of linguistic models is effective in improving many tasks related to the processing of natural languages, including the intuition and paraphrase of natural languages, the recognition of named entities and, the answer to questions. The development of pre-trained language models based on transformer architectures has stimulated the evolution of modern techniques for many tasks in the field of natural language processing (NLP) [20,21,22]. The study of [23] proposed text classification using BERT for natural language processing and the results of the experiment showed that combinations of BERT with CNN, RNN, and BiLSTM performed well with precision, recall rate, and F1 score, compared to Word2vec. The new BSTC (BERT, SKEP, and TextCNN) fake review detection model is proposed [24] based on a pre-trained language model and a convolutional neural network. The highest accuracy was achieved with all three gold standard datasets (Hotel, Restaurant, and Doctor), with 93.44%, 91.25%, and 92.86%, respectively. The process of choosing, modifying, and transforming raw data into features that can be utilized to enhance the performance of machine learning models is known as feature engineering. In some tasks, effective feature engineering combined with conventional machine learning methods could produce outcomes comparable to BERT [25]. Although there has been a rise in interest in learning general-purpose sentence representations, the majority of the research in that field has been conducted in English and has mostly been monolingual [26].

Spam is typically defined as undesired text that is sent or received over social media platforms like Facebook, Twitter, YouTube, e-mail, etc. [27]. The authors of [28] proposed a novel four-layered, state-of-the-art detection strategy, with graph-based, neighbor-based, automation-based, and time-based features to find spammers on social networking sites. The majority of SMS spam classifiers use supervised algorithms like Naïve Bayes (NB), support vector machine (SVM), neural networks, and regression, because the availability of the output column (labeled data) of the SMS dataset makes it possible to train classification models [29]. Using a total of 20 samples from the dataset (SMS Spam Corpora and Twitter Corpora), the suggested solution in [30] employs reinforcement learning to identify the malicious social bots. It also makes use of k-nearest neighbor (KNN) and a recurrent neural network (RNN). A social bot is a computer program that uses an application programming interface (API) to operate a social media account. It can be used for malicious activities, such internet trolling and fraud. Bots are classified as malicious or benign in the study cited [31].

Information phishing began as a marketing tactic, but it has since evolved into destructive internet interactions that expose users to significant security risks using tools including emails, comments, blogs, and messaging. Given their adaptability and ability to make the most of current hardware and computational limitations, deep learning architectures like convolutional neural networks (CNNs), multi-layer perceptrons (MLPs), and the long short-term memory (LSTM) have been successfully used for email spam classification [32]. The identification of fake news [33,34] is a difficult challenge for social media platforms like Facebook, Twitter, etc., because of the volume of data that people publish on these sites. To determine whether a news article is authentic or fake, a deep CNN for fake news detection was presented in [35] and models were tested using binary class datasets. For NLP researchers, sarcasm presents a formidable challenge and can entirely alter the meaning of a statement, making it challenging for modern models and systems to recognize it. In order to create models that can accurately identify the settings in which sarcasm may occur or is suitable, an approach for the automatic detection of sarcasm context has been developed [36].

Cyber social media security examines the dynamics of online social networks, the data’s vulnerabilities, and the potential effects of their abuse by social media attackers. Due to their nature, the volume of content they include, and the sensitive information they use, social media are the most attack-prone section of the internet [37,38,39]. To classify a social media message as a part of a particular crisis event, it is important to take into account a number of factors, such as the message’s nature, the information it contains, the source of that information, its credibility, the timing, and its location [40]. Some of the features can be automatically extracted, whereas some need to be manually labeled. The best performance is achieved with an ensemble approach for the identification and classification of crime-related tweets that uses logistic regression (LR), SVMs, KNN, a decision tree (DT), and an RF classifier assigned the weights of 1, 2, 1, and 1, respectively, ensemble together via a soft weighted voting classifier along with a term frequency–inverse document frequency (TF-IDF) vectorizer with an accuracy of 96.2% on the testing dataset [41]. When compared to the ground truth labeled by network experts, an RNN-LSTM model that was trained to identify five different social engineering attacks (SEA) that may show signs of information gathering achieves classification precision and recall scores of 0.84 and 0.81, respectively [42].