Optimized Deep Learning Techniques for the Detection and

Identification of Fake News in Digital Media

Narmadha Devi A. S.

, K. Sivakumar

and V. Sheeja Kumari

Department of Engineering Mathematics, Saveetha School of Engineering, Saveetha Institute of Medical and Technical

Sciences, Saveetha University, Chennai, Tamil Nadu, India

Department of Computational Intelligence, Saveetha School of Engineering, Saveetha Institute of Medical and Technical

Sciences, Saveetha University, Chennai, Tamil Nadu, India

Keywords: Deep Learning, Long Short‑Term Memory (LSTM), N‑Gram, Porter's Stemming, Social Networks, Fake

News Detection.

Abstract: The exponential expansion of social media has greatly accelerated the dissemination of disinformation,

endangering public safety and undermining faith in news outlets and government agencies. The authors of

this work suggest using deep learning to identify false news posts on Twitter. The methodology involves pre-

processing raw data through stop word removal, stemming using Porter’s Algorithm, and tokenization with

the N-gram model. The detection model employs Long Short-Term Memory (LSTM), Convolutional Neural

Networks (CNN), and AdaBoost algorithms. Results indicate that LSTM outperforms CNN and AdaBoost,

achieving an accuracy of 99.24%, specificity of 99.2%, and sensitivity of 98.67% in fake news detection.

1 INTRODUCTION

Fake news spreads misinformation, misleading

people and impacting society, businesses, and

individuals. It harms reputations, leaving lasting

damage even after corrections. By reinforcing biases,

it deepens divisions, fosters distrust, and fuels

conflicts. Politically, misinformation manipulates

public opinion to serve specific agendas.

Additionally, the rise of fake news erodes trust in

journalism, weakening democracy and

accountability. Greater emphasis has been placed on

identifying and removing disingenuous material due

to the spread of false information on social media

platforms like Twitter and Facebook. Misleading

information spreads rapidly, often diverting attention

from critical issues, with many users trusting social

media over traditional sources despite scepticism

about reliability (t’Serstevens,et al, 2022).

Confirmation bias further reinforces misinformation,

making deception harder to recognize. False news

appears in various formats, including articles, images,

and videos, contributing to widespread confusion.

Rumours can disrupt social harmony and cause

significant societal impact. Additionally, fake

websites mimic credible sources to manipulate public

perception, influencing opinions and advancing

political or financial agendas (Igwebuike, E. E., &

Chimuanya, L. 2021). A complicated process for false

information detection is depicted in Figure 1.

Figure 1: Block diagram of fake news detection.

The original data is preprocessed by gathering

tweets from Twitter using aggregation. The dataset is

passed through a preprocessing phase, which includes

the removal of stop words, stemming, and

tokenization. Features for the classification model are

identified to make it more efficient. Machine learning

and deep learning methods are then used to classify

tokenized data. As a result, techniques from deep

506

S., N. D. A., Sivakumar, K. and Kumari, V. S.

Optimized Deep Learning Techniques for the Detection and Identiﬁcation of Fake News in Digital Media.

DOI: 10.5220/0013932100004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 5, pages

506-511

ISBN: 978-989-758-777-1

learning and machine learning are used to identify

rumours and false news.

The advent of online and social media has enabled

the incorporation of false information with real or

verified information. This situation can be utilized to

influence people’s opinions, thus impacting their

perceptions, thoughts, and behavior. As a result,

disseminating links, messages, photos, videos, and

audio files over several social media platforms has

become very simple for those who propagate fake

news. People who spread these fakes usually have a

political or social agenda. Therefore, the development

of an efficient system to detect misinformation is of

utmost importance (Kaliyar, R. K., et al, 2021). An

approach to detecting false news stories using deep

learning is presented in this research. There are input

datasets that make up methodology. Information

culled from the microblogging service Twitter is the

source of this dataset. Input data that is in its raw form

undergoes data preparation initially. Remove Stop

Words, Stemming, and Tokenization are the main

components of data preparation. Use the NTLK

library to remove stop words. Stemming is done using

Porters Algorithm. Tokenization is completed by N-

gram model. Model is developed using LSTM, CNN

and AdaBoost algorithms. Results have shown that

LSTM’s Compared to CNN and AdaBoost methods,

the accuracy, specificity, and sensitivity are higher.

2 LITERATURE REVIEW

In order to identify false news, researchers use n-gram

analysis and TF-IDF for feature extraction. After that,

they use decision trees, SGD, Linear SVM, Logistic

Regression, SVM, and KNN as machine learning

classifiers (Lahby, M.,et al, 2022). Pennycook & Rand

(Pennycook, G., & Rand, D. G. 2021).

developed an

SVM-based satire detection model with 90%

accuracy. Bahad et al. showed RNNs outperform

manual rumor detection, while Ruchansky et al.

introduced the CSI model, integrating content, user

comments, and sources for improved accuracy. For

fake images, Hsu et al. developed CFFN, using GANs

and DenseNet to classify manipulated images. Bird et

al. developed NLTK, a comprehensive Python toolkit

that facilitates various NLP tasks such as

tokenization, parsing, stemming, and classification,

making text analysis more accessible and efficient.

Huan et alsuggested a deep learning strategy for text

classification that effectively captures both sentiment

and semantic context, improving accuracy in

emotionally charged text analysis (Bird, S., Klein, E.,

& Loper, E. 2009)

Umer et al. demonstrated that

combining convolutional neural networks (CNNs)

with Fast Text embeddings enhances text

classification by efficiently extracting contextual and

syntactic features (Umer, M.,et al, 2023). Optimized

deep learning methods for spotting rumors and

misleading information in online social networks

were presented by Zamani et al, leveraging advanced

neural architectures to enhance misinformation

identification and content credibility assessment

(Abu Sarwar Zamani, et al, 2025).

3 RESEARCH METHODOLOGY

3.1 Deep Learning for Fake News

Detection

This technique involves collecting data from Twitter

in order to utilize deep learning to identify false news.

Pre-processing involves stop word removal (NLTK),

stemming (Porter’s Algorithm), and tokenization (N-

gram model). Tokenization applies unigrams,

bigrams, and trigrams to structure text. The model

integrates LSTM, CNN, and AdaBoost for

classification. LSTM, an RNN variant, is effective in

pattern recognition due to its input (I/P), forget (f),

and output (O/P) gates, along with a memory cell.

These gates regulate information flow, ensuring data

integrity and sequence retention, improving accuracy

while preventing gradient descent issues. Figure 2

and 3 illustrate the LSTM network, providing A

schematic illustration of its composition. The network

takes an embedding xi as input at each time step and

calculates its output hi by adding the output h i-1

together with the latest embedding xi to the latest cell

state h i-1. It is possible to insert or remove data from

the cell, depending on its present state

Figure 2: LSTM architecture.

Optimized Deep Learning Techniques for the Detection and Identiﬁcation of Fake News in Digital Media

507

𝑓𝑡 = 𝜎(𝑤𝑓. 𝑦𝑡 − 1, 𝑥𝑡 + 𝑏)

(1)

𝑖𝑡 = 𝜎(𝑤𝑖. 𝑦𝑡 − 1, 𝑥𝑡 + 𝑏)

(2)

A Convolutional Neural Network (CNN) comprises

key components like convolution and pooling layers.

While CNNs are well-known for image processing, they

also identify data interdependencies. Feature extraction

from input data is made possible by the convolution

layer, which allows operations on the embedding matrix

for word embeddings. The pooling layer then reduces

dimensionality and selects important features using

methods like max, min, or average pooling. Finally, the

extracted features are processed by a fully connected

neural network. A CNN applies activation functions to

generate the final output, typically consisting of

convolutional, pooling, activation, and fully connected

layers. Deep CNNs enhance learning by stacking

multiple convolutional layers, which act as filters,

processing small pixel sections at a time (e.g., 3×3

filters). Researchers at the University of Michigan

developed AdaBoost, an advanced gradient-boosting

method for binary classification. It starts with an initial

decision tree, evaluates its accuracy, and integrates

multiple classifiers to build a robust model.

The first model is built with the training data and is

then enhanced by including other models to address its

shortcomings. This process continues until all training

data is accurately predicted or the model limit is reached.

To enhance accuracy, multiple classifiers are combined

into a single optimized model. AdaBoost is widely used

for pedestrian detection, where images are cropped into

sections, and marking windows help identify pedestrians.

The same method is applied iteratively with different

selection sequences.

Figure 3 shows Convolutional Neural

Network.

A window is classified as containing a

pedestrian if no models reject it, refining the

classification process further.

Figure 3: Convolutional neural network.

3.2 Dataset Overview and

Preprocessing

The dataset, sourced from Kaggle, contains 40,000

articles 20,000 real and 20,000 fake. A pre-trained Glove

Twitter dataset is also referenced.

Figure 4: Present tendencies across different types of fake

news and real news.

Data pre-processing is a crucial step to enhance

efficiency, involving data transformation and cleaning

before execution. The following section elaborates on

these processes. A figure illustrates that balanced,

indicating stability. In the visualization, the "0" class

(orange bar) represents fake news, while the "1" class

(blue bar) represents real news. Since topic contents vary

between categories, only the main text is processed,

while the subject, title, and date are removed. Figure 4

explains the topics that comprise the dataset, showing

how news spreads in society. The count of each subject

highlights its presence in the dataset—blue bars denote

unreliable news, while orange bars indicate credible

news. Real news covers global and political topics,

whereas fake news is often found in categories like

politics, general news, left-wing media, U.S. news, and

Middle Eastern news.

3.3 Deep Learning LSTM Model

A sequential model has one input and one output tensor

per layer, requiring a defined input shape. Sequence

classification predicts categories based on sequential

data, posing challenges due to varying lengths and

complex patterns. To detect false information, an LSTM

model is used, leveraging LSTM cells with input, forget,

and output gates. Figure 2 shows how a tan layer

regulates the combination of the previous output ht−1

with the value of the new sequence, xt in order to smooth

inputs.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

508

3.3.1 LSTM Cell Process

The LSTM model begins with a tanh layer that

smooths the combined input. Next, the input gate,

with sigmoid activation, filters relevant values by

scaling the compressed input. The forget gate (st)

determines which information to retain, while the

previous state (st−1) is added to maintain long-term

dependencies, reducing the risk of gradient

disappearance. The forget gate further refines stored

information, ensuring only necessary data is

preserved. Finally, the output gate regulates the final

output using a tanh function, determining which

values from the cell (ht) are allowed as output.

The input is reduced to a range of -1 to 1 by using

a tanh activation function. Shown below is one way

to do it:

𝑔=𝑡𝑎𝑛ℎ(𝑏



+𝑋



𝑈



+ℎ



𝑉



)

(3)

where Vg stands for the input weight and bg for the

preceding cell output, with bg representing the input

bias.

The symbol for the forgotten gate is:

𝑓=𝜎(𝑏



+𝑥



𝑈



+ℎ



𝑉



(4)

The output is st−1, which is the element-wise sum of

the prior state and the forget gate.

It is possible, nonetheless, to represent output gates

0=𝜎(𝑏



+𝑥



𝑈



+ℎ



𝑉



) (5)

As shown in Figure 2, the final output of the

network is ht. To enhance performance, a stacked

LSTM model is used with the return sequence set to

true, allowing each neuron’s hidden state to serve as

input for the next LSTM layer. Two long short-term

memory (LSTM) layers, one with 128 memory units

and the other with 64, follow each word is represented

as a 32-length vector in the embedding layer model.

The first dense layer activates 32 memory units and

uses Re LU, while the second layer uses a sigmoid

function to train a single neuron for output. Neurons

in a dense layer are completely connected, so data

from all neurons in the layer below it may reach each

one.

To prevent overfitting, recommended networks

use one or two dense layers. ReLU, a widely used

activation function, prevents simultaneous neuron

activation and offers advantages over sigmoid and

tanh in convolutional neural networks. ReLu is

represented by:

𝜎=𝑚𝑎𝑥(0, 𝑧) (6)

If Qij is sigmoid, then the likelihood of the word j

occurring in connection to the term i is this. This is

how we may express the inferred global objective

function:

𝐽 = −∑ LogOiji ∈ corpus, j ∈ context(i) (7)

A Dense output layer may be used to classify false

news as either legitimate news (with a value of 0) or fake

news (with a value of 1). Utilizing the optimizer, metrics,

and loss function during model construction is essential.

Ten iterations of training the model using the Binary

Cross-Entropy loss function and a learning rate of 0.01

are implemented using the Adam optimizer. Decreased

batch size, set at 256, has improved accuracy. The size

of the embedding is 100. A random sample strategy was

used in the execution of the investigation. From one of

Bhutan's Colleges of Education, 22 first-year in-service

postgraduate science teachers made up the sample.

Males made up 13 (or 59% of the total) of these

educators. Because of this, picking the right

hyperparameters for a model is crucial for fast and

accurate training. Internal operations of the cells that

make up the LSTM network are the main point of

differentiation.

Table 1: LSTM Layered Architecture.

Layer (type) Output size

Param

Numbe

Embedding_1

(embedding)

300 x 100 10,00,000

Lstm_1 (LSTM) 300 x 128 1,17,248

Lstm_2 (LSTM) 64 49,408

Dense_1 (Dense) 32 2080

Dense_2 (Dense) 1 33

Table 2: Hyperparameters for proposed model.

Hyperparameters Value

Layer for embedding 1

LSTM layer 2

Layer with high concentration 2

Loss Function

Binary cross

entrop

Function for activation ReLu

Optimizer Adam

Learning rate 0.01

Epoch count 10

Size of embedding 100

Group quantity 256

Optimized Deep Learning Techniques for the Detection and Identiﬁcation of Fake News in Digital Media

509

Tables 1 and 2 provide a comprehensive list of all

the necessary hyperparameters for an LSTM model to

improve performance, as well as the recommended

settings considered best practices.

4 RESULT AND DISCUSSION

The dataset includes four attributes title, main text, topic,

and date and is derived from Twitter. It includes pre-

trained word vectors, with 20,000 features used for

analysis 16,000 for training, 2,000 for testing, and 2,000

for validation. Vectorization is based on word frequency.

Figure 5 shows for a comparison of classifier results in

detecting false news in social media datasets. Data pre-

processing involves removing stop words using the

NLTK library, stemming words with Porter’s

Algorithm, and tokenizing with an N-gram model.

Convolutional neural networks (CNNs), boosted by

AdaBoost and LSTM, form the basis of the model.

LSTM achieved 99.24% accuracy, outperforming CNN

by 1.68% and AdaBoost by 5.02%. Its specificity is

99.2%, exceeding CNN by 2.04% and AdaBoost by

4.99%, while its sensitivity is 98.67%. LSTM is the most

effective for fake news detection.

Table 3: Accuracy, specificity and sensitivity comparison

of different classifiers.

Algorithm/Metric

Accuracy

(%)

Sensitivity

(%)

Specificity

(%)

AdaBoost 94.22 96.55 94.21

CNN 97.56 96.33 97.16

LSTM 99.24 98.67 99.2

Figure 5: Comparison of classifier results in detecting false

news in social media datasets.

Additionally, rhetoric plays a key role in English writing

by enhancing persuasive abilities. Table 3 shows

accuracy, specificity and sensitivity comparison of

different classifiers. Understanding rhetorical devices

like contrast and exaggeration helps writers improve

their skills and grasp rhetorical concepts more

effectively.

5 CONCLUSIONS

This research presents a method for detecting false news

stories using Twitter data that is based on deep learning.

Tokenization, stemming, and stop word removal are all

part of the pre-processing. When compared to CNN and

AdaBoost, LSTM achieves the highest accuracy rate of

99.24%. Future work aims to enhance automation,

particularly for e-commerce platforms, where detecting

false information is crucial.

REFERENCES

Abu Sarwar Zamani, Aisha Hassan Abdalla Hashim, Sara

Saadeldeen Ibrahim mohamed, Md. Nasre alam.

Optimized Deep Learning Techniques to Identify

Rumors and Fake News in Online Social Networks.

Journal of Computational and Cognitive Engineering

2025,Vol.00(00)19DOI:10.47852/bonviewJCCE52023

348.

Bahad, P., Saxena, P., & Kamal, R. (2019). Fake news

detection using bi-directional LSTM-recurrent neural

network. Procedia Computer Science, 165, 74–82.

https://doi.org/10.1016/j.procs.2020.01.072

Bird, S., Klein, E., & Loper, E. (2009). Natural language

processing with Python: Analyzing text with the natural

language toolkit. USA: O’Reilly Media, Inc.

Hsu, C. C., Zhuang, Y. X., & Lee, C. Y. (2020). Deep fake

image detection based on pairwise learning. Applied

Sciences,10(1),370.https://doi.org/10.3390/APP10010

370

Huan, J. L., Sekh, A. A., Quek, C., & Prasad, D. K. (2022).

Emotionally charged text classification with deep

learning and sentiment semantic. Neural Computing

and

Applications,34(3),23412351.https://doi.org/10.1007/s

00521-021-06542-1

Igwebuike, E. E., & Chimuanya, L. (2021). Legitimating

false hood in social media: A discourse analysis of

political fake news. Discourse & Communication,

15(1), 42–58. https://doi.

org/10.1177/175048132096165

Kaliyar, R. K., Goswami, A., & Narang, P. (2021).

DeepFakE: Improving fake news detection using tensor

decomposition- based deep neural network. The Journal

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

510

Supercomputing,77(2),10151037.https://doi.org/10.10

07/s11227-020-03294-y

Lahby, M., Aqil, S., Yafooz, W. M. S., & Abakarim, Y.

(2022). Online fake news detection using machine

learning techniques: A systematic mapping study. In M.

Lahby, A. S. K. Pathan, Y. Maleh & W. M. S. Yafooz

(Eds.), Combating fake news with computational

intelligence techniques (pp.

337).Springer.https://doi.org/10.1007/978-3-030-

90087-8_1

Pennycook, G., & Rand, D. G. (2021). The psychology of

fake news. Trends in Cognitive Sciences, 25(5), 388–

402. https://doi. org/10.1016/j.tics.2021.02.007

Ruchansky, N., Seo, S., & Liu, Y. (2017). Csi: A hybrid

deep model for fake news detection. In Proceedings of

the 2017 ACM on Conference on Information and

Knowledge

Management,797806.https://doi.org/10.1145/3132847.

3132877

t’Serstevens, F., Piccillo, G., & Grigoriev, A. (2022). Fake

news zealots: Effect of perception of news on online

sharing behavior. Frontiers in Psychology, 13, 859534.

https://doi.org/ 10.3389/fpsyg.2022.859534.

Umer, M., Imtiaz, Z., Ahmad, M., Nappi, M., Medaglia, C.,

Choi, G. S., & Mehmood, A. (2023). Impact of

convolutional neural network and FastText embedding

on text classification. Multimedia Tools and

Applications, 82(4), 5569–5585. https://

doi.org/10.1007/s11042-022-13459-x

Optimized Deep Learning Techniques for the Detection and Identiﬁcation of Fake News in Digital Media

511