Home / << Go Back
📘 ICNERDP02
Authors: Amanuel Bekele¹, Selamawit Mekonnen², Dawit Tesfaye³
Department: International Conference Nerd Publication
Volume:
Vol 1 / Issue 1 (2019)
Submitted: 28 Jul 2025
content: Assessing Adversarial Vulnerabilities in
Fake News Detection: A Comparative Study
of GPT-2 and BERT Variants
Ishan Bhardwaj¹, Vaibhav Agrawal², Vaibhav Pratap Singh³, Dinesh Kumar Vishwakarma⁴
1,2,3Department of Information Technology, Delhi Technological University, Delhi, India
⁴Professor, Department of Information Technology, Delhi Technological University, Delhi, India
1
theishanbh@gmail.com, 2vaibhavagr514@gmail.com, 3vaibhav14112002@gmail.com,
4dvishwakarma@gmail.com
Abstract: The increasing sophistication of fake news dissemination poses a growing threat to digital
information integrity, demanding the deployment of robust and intelligent detection systems. Transformer
based language models—particularly BERT, RoBERTa, DistilBERT, and GPT-2—have shown promising
results in detecting misinformation by leveraging deep contextual understanding. However, their vulnerability
to adversarial attacks reveals a critical weakness in their deployment for real-world applications. This study
conducts a comprehensive evaluation of these models under both standard and custom adversarial attack
scenarios to assess their reliability in detecting manipulated or misleading content. Using the
"newsmediabias/fake_news_elections_labelled_data" dataset, we fine-tune each model and subject them to a
battery of adversarial techniques, including TextFooler, PWWS, BAE, DeepWordBug, TextBugger, as well as
novel attack methods designed specifically for this study: Enhanced Substitution Attack (ESA) and
Comprehensive Text Attack (CTA). We analyze model behavior in terms of accuracy degradation, perturbation
efficiency, and computational cost. Our findings reveal stark contrasts in model robustness: while RoBERTa
maintains the highest performance on clean data, it—along with other models—is significantly compromised
under even subtle adversarial manipulations. The study highlights GPT-2's limitations as a generative model
repurposed for classification, as it fails catastrophically under most attack conditions. These insights
underscore the urgent need for adversarial resilience in fake news detection systems and pave the way for
future research focused on integrating robust defense mechanisms into transformer-based architectures.
Keywords: Adversarial Attacks, Fake News Detection, BERT, RoBERTa, DistilBERT, GPT-2, TextAttack,
Model Robustness, NLP Security
1.
INTRODUCTION
The rapid proliferation of misinformation and disinformation through digital platforms poses a
significant threat to democratic institutions, public health, and societal trust. The evolution of algorithmic content
recommendation systems has amplified the scale and speed at which fake news spreads, intensifying the need for
automated and robust detection frameworks. In this context, machine learning models—particularly those rooted
in natural language processing (NLP)—have become indispensable. Transformer-based architectures such as
BERT (Bidirectional Encoder Representations from Transformers) [2], RoBERTa [3], DistilBERT [4], and GPT
2 [5] have redefined the state-of-the-art in text classification tasks, including sentiment analysis, fact verification,
and fake news detection, due to their capacity to capture contextual semantics and syntactic structures [1, 10].
Early detection efforts primarily employed rule-based or statistical methods, which proved insufficient
for handling the nuanced semantics of natural language. The introduction of the Transformer architecture [1] and
its successors—like BERT and RoBERTa—enabled bidirectional contextual understanding, resulting in marked
improvements in classification accuracy. DistilBERT further optimized BERT’s capabilities for real-time
applications through knowledge distillation, while GPT-2, a decoder-only model designed for generative tasks,
has also been adapted for classification using techniques such as LoRA (Low-Rank Adaptation) [36].
Despite these advancements, recent research has highlighted a critical limitation: transformer models are
highly vulnerable to adversarial attacks. These attacks involve subtle, often imperceptible perturbations to input
text that preserve human readability but mislead model predictions [11, 13, 15]. Such adversarial manipulations
are especially dangerous in politically sensitive or high-stakes domains, where misinformation can influence
public opinion or policy. Frameworks like TextAttack [12] have enabled the systematic generation of these
adversarial examples using various algorithms—TextFooler [8], PWWS [13], DeepWordBug [14], BAE [31], and
TextBugger [15]—which target both word-level semantics and character-level structures. Research shows that
these methods can reduce model performance by over 90%, even with minor textual alterations [11, 15, 35].
Encoder-based models like BERT and RoBERTa generally perform well under clean testing conditions
but suffer notable performance degradation under adversarial stress [16, 33]. Decoder-only models such as GPT
www.ijama.in
P age | 15
International Journal of Advanced Multidisciplinary Application | IJAMA
Volume 2 Issue 5 May 2025
ISSN No: 3048-9350
2 exhibit even greater vulnerability due to their unidirectional architecture and lack of specialized objectives for
discriminative tasks [30, 35]. While defenses such as adversarial training [17], input preprocessing [48], and
parameter-efficient fine-tuning methods like LoRA [36, 37] have been proposed, they offer only partial mitigation
and are often underexplored in the context of fake news detection.
To address these challenges, this study conducts a comprehensive comparative analysis of four
📘 ICNERDP01
Authors: Prof.NERD ,Prof.Madan
Department: International Conference Nerd Publication
Volume:
Vol 1 / Issue 1 (2019)
Submitted: 28 Jul 2025
content: Summation Formulae for a Series Involving
the I – Functions of Several Variables
P C Sreenivas1, T M Vidya2, T M Vasudevan Nambisan3, P V Maya4
1Principal, Gurudev Arts and Science College Mathil, Payyanur, 2Assistant Professor, Department of Mathematics, Mahatma
Gandhi College Iritty, Kannur, Kerala, 3Emeritus Professor, College of Engineering Trikaripur, Kasargod, 4Assistant Professor,
Department of Mathematics, Mahatma Gandhi College Iritty, Kannur, Kerala
1sreenivaspc@rediffmail.com, 2vidyatm4@gmail.com, 3tonambisan.tm@gmail.com, 4panakkalveettilmaya@gmail.com
Abstract- In this paper, we derive two new summation formulae for a class of series involving I – functions of
several variables. These functions, which generalize many known special functions, play a crucial role in
various branches of Mathematical analysis and applied Mathematics. By employing suitable techniques of
summation and transformation, we establish compact and elegant expressions for these series under certain
convergence conditions. These results contribute to the ongoing development of summation theorems for
generalized special functions and may further applications in Mathematical analysis.
Keywords: I – function of several variables, Gamma function, Special functions, Pocchammer symbol