Table 1 Comparison between various methods to identify FNs

[12]

ISOT and FA-KES

The study introduced a novel hybrid DL approach integrating RNNs and CNNs for FND

The approach requires combination with task-specific feature engineering methods to be useful

The proposed model introduced the concept of generalized methods in handling FND

The suggested approach exceeds other non-hybrid approached in terms of classification accuracy

[13]

More than 15 thousands of news from different users of the Facebook containing both truthful news and FNs

Incorporating the behaviour of various attributes linked to Facebook accounts and examining those accounts behaviour utilizing DL and ML methods.

DL methods need more time to train and test than Ml methods

Check the features of both user content and news content to identify FNs on Facebook

Accuracy LR = 99.0, DT = 99.1, k-NN = 99.3, SVM = 99.3, LSTM = 99.4

[40]

FNs data collected by Jruvika and FNs data collected by Guilherme Pontes

A novel CNNs semi- supervised model depend on the self-ensembling method to utilize the benefit of the linguistic information and stylometric of annotated news articles and identify the hidden patterns in unlabelled data

The proposed methodology didn’t utilize transformers to enhance the accuracy of classification

ConvNet filters are employed separately on headline part and body of the news articles and then chosen feature vectors were integrated to take benefit of both the slices

The proposed approach obtained 97.45% accuracy of classification on FNs data collected by Jruvika dataset

[26]

LIAR

The study presented an efficient DL approach that utilized to determine the fakeness level in the reports of news

The accuracy of classification is not effective

The proposed algorithm integrated the utilize of attention mechanisms with relevant metadata available and uses contextual embedding as word embedding

The proposed approach yielded 46.36% accuracy of classification on the LIAR dataset, which exceeded others by 1.49%

[41]

Collected 1356 records of news from multiple users via media sources such as Twitter and PolitiFact and create different datasets for the FNs and truthful news

They combined bidirectional-LSTM with CNNs networks with attention mechanism to generate more efficient accuracy after transcribing text into vectors using 100-dimensional GLOVE word embedding

The proposed approach didn’t result in a more reliable classification accuracy

The proposed methodology utilized two multiple techniques

Bidirectional-LSTM with CNNs integrated network with recognition technique achieved the highest classification accuracy of 88.78%

[42]

Two real-world datasets in two languages ( English and Korean)

A novel embedding technique named link2vec extended from word2vec

The performance of the suggested link2vec should be checked via various search engines such as Google

The proposed link2vec performance is greater than that of text based methods in the two language FNs datasets

The proposed methodology outperformed text based techniques in the two utilized datasets

[43]

Aggregate real datasets from Twitter

They compared different types of ML and DL methods to discover FNs intweets, aiming at identifying patterns in both linguistic content and structure of tweets

They did not investigate the significance of image information

Combining conventional ML and DL techniques improved proposed model behaviour through the latter, while also obtaining insight towards tweets structure from the interpretability of the former

DL techniques outperformed the conventional methods, getting 99% F1 score

[16]

The dataset relates to FNs prevalence during the United States presidential election in 2016

The proposed methodology is introduced to automatically recognize the discriminating properties to classify FNs through multiple hidden layers constructed in the DNN

They didn’t utilize transformers that can increase the accuracy of classification

The proposed approach will assist researchers to extend understand of the CNN-based deep models applications to detect FNs

Obtained classification accuracy of 98.36%

[44]

Buzzfeed corpus, SFL dataset, FND dataset, and satire political dataset

Introduced ensemble learning framework combining four different methods called embedding depth LSTM , LSTM, LIWC CNN, and N-gram CNN for detecting FNs.

They need to investigate grammatical analysis in depth

Create optimized weights , improve the precision rate, and investigate the intractability problem through different domains

They achieved classification accuracy of 99.4%

[45]

Twitter (shared news and user profiles)

Describe a deep analysis of the characteristics which from a human and and automation views are more predictive to identify social network profiles which distributes FNs in the online environment.

The proposed model need to be verified on multiple datasets of news and users to test the accuracy of prediction

Identify the trusted profiles. Verify the reliability of social information and distributed articles.

Displaying better information enables humans and machines to determine malicious users, the average classification accuracy of 90%

[46]

COVID-19, LIAR, and ISOT

The proposed model integrated multi-layer perceptron, single layer perceptron, and CNN after the embedding layer which including pre-trained approached such as BERT, RoBERTa, GPT2, and Funnel Transformer to benefit from the deep contextual representation produced by those models and classifications ability of deep neural models

They need to use user profiles for more information

Obtaining positive outcomes without using extra attributes and network data improves the capability of learning, making contextual-dense embedding of input texts

The empirical outcomes show enhancements in the classification accuracy in comparison with traditional approaches in the same datasets

[47]

FNs or truthful news dataset, LIAR , and corpus both truthful news and FNs

Proposed the pre-trained methods for detecting FNs along with conventional and DL approaches and compare their results from different views

They need to construct health related FNs methods that will be deployed in social media during the COVID-19 pandemic

Their findings will help the research community to investigate more and news sites to select the most appropriate approach for detecting FNs

Pre-trained methods behave well to detect FNs, specially with small-scale datasets