AI frameworks for fake news detection

The rise of social media has transformed the way information spreads, making fake news detection a critical challenge. To address this issue, Dr. Fadi Mohsen, a UG researcher specializing in cybersecurity and machine learning, has developed a web framework that enables researchers to compare various machine learning algorithms for misinformation detection.

Dr. Mohsen’s initial research focus was on machine learning, but his fascination with the ever-evolving nature of security threats led him to cybersecurity. During his PhD, he identified vulnerabilities in Android systems and authentication protocols, working to develop countermeasures and improve security practices. Over the past nine years, his expertise has expanded. While cybersecurity remains his primary focus, he has consistently integrated machine learning, text mining, and large language models (LLMs) into his research.

In this interview, we explore his latest dataset and web framework, ‘Machine Learning Frameworks for Fake News Detection and Datasets’, designed to help researchers compare machine learning models for fake news detection.

Could you provide an overview of your dataset and its significance in the context of fake news detection?

In this research, we utilized multiple datasets that have been previously referenced in the literature. To better align with real-world conditions, particularly for incremental learning scenarios, we derived new datasets from existing sources, including trimmed versions of WELFake, Scraped, and Kaggle 1.

To promote reproducibility and further research, we have made the original datasets, newly generated datasets, and the scripts used for dataset generation publicly available. This ensures that other researchers can replicate our experiments, validate our findings, or extend our approach by applying it to new contexts. Our work provides a structured framework for evaluating fake news detection models across evolving datasets, which is crucial for tackling misinformation in dynamic environments.

You published your code along with your data. Did you face any challenges in preparing the code for publication?

Yes, we built a framework that allows researchers to conduct comparative studies on mainstream machine learning algorithms and their ability to detect fake news articles. One of the challenges we faced was writing a clear and detailed README file to ensure that others could easily understand and use our framework. While our current method of sharing the code is sufficient, we realized that it lacks proper version control for tracking changes made by others.

You have chosen to publish your code and dataset in the repository DataverseNL. To what extent can open access to research data help to counter misinformation and fake news?

Open research data and code are crucial in every field, including fake news detection. The spread of misinformation can cause panic, uncertainty, and even real-world harm. By adopting an open science approach, researchers can collaborate more effectively, validate findings, and accelerate advancements in misinformation detection. Publishing datasets and code in open repositories like DataverseNL ensures Transparency & Reproducibility, Efficiency & Progress, and Stronger Countermeasures.

One of the big worries with AI is that we are unconsciously introducing biases in the algorithms. Did you identify any biases in your code? 

We utilize AI and machine learning in crafting cybersecurity solutions. Although we tend to tackle the possible biases introduced by the selected algorithms. Our approach does not dive deep into the core implementations of these algorithms; instead we focus on the experimentation settings to limit any possible biases. For example, in our fake news detection work, we purposely selected various and diverse datasets to mitigate potential biases and enhance the generalizability of our findings.

Given your expertise in the field of fake news detection, what is the most striking thing you have found in your work?

There are two things that strike me the most: first, the volume and the widespread use of fake content on the web. Second, the utilization of  AI and large language models (LLMs) by cyber criminals to generate more believable and hard to detect fake news articles, images, and videos. 

On the one hand, the knowledge and the technical capability to deal with fake news detection is evolving, while on the other hand, societies remain greatly vulnerable to the dangers of fake news and misinformation.  What macro-level strategies should be adopted to better protect societies from this global challenge?

In the world of security, humans are regarded as the weakest link. The root cause of vulnerabilities in systems are human malpractices e.g. a developer forgets to sanitize data before sending them to the backend or a system admin uses easy passwords. Furthermore, the exploitation of these vulnerabilities also occurs due to the human’s error, e.g., an employee clicking on an email attachment from a strange address or an employee forgetting to change her password. The same can be applied to the spread of fake news. Users have the tendency to share information without verifying it. Therefore, advancing the automated countermeasures is not enough to counter this volume of misinformation. Educating the public on the need to verify before sharing any information is crucial.