Deluminator

Inspiration

In today's digital world, misinformation spreads rapidly, shaping opinions and decisions. Inspired by the need for trustworthy information, we developed a browser extension that detects and corrects fake news using machine learning and NLP.

What it does

We built a chrome extension that detects the likelihood possibility of the contents of the active website being fake news using BERT from huggingface, while parsing the text to find the top 5 most repeated terms. The extension then searches the web with these terms and returns the top 3 most relevant results as links, allowing for the user to easily cross reference these 3 other sources, to have a better understanding of the matter and not fall for fake news.

How we built it

We trained our model using the Fakeddit multimodal fake news classification dataset, applying text preprocessing techniques such as tokenization, stop-word removal, and lemmatization. We started with baseline models like Simple Regression for quick validation. We then move on to Binary Classificaiton. Used Early stopping to minise loss, AdamW, CrossEntropyLoss (Binary) / BCEWithLogitsLoss (Regression) for fine-tuning

Challenges we ran into

We faced several challenges, including data imbalance (more real than fake news), highly deceptive articles, and scalability issues when deploying BERT in a browser extension.

Accomplishments that we're proud of

A key feature of our extension is a credibility scale, providing users with a confidence score on the likelihood of misinformation.

What we learned

Throughout this project, we gained valuable insights into Natural Language Processing (NLP) and Large Language Models (LLMs), including text analysis, tokenization, and feature extraction. We explored machine learning models, from traditional classifiers like Logistic Regression to advanced Transformer-based models like BERT. Additionally, we focused on data preprocessing, including cleaning, normalization, and vectorization to enhance model performance. Lastly, we addressed ethical considerations, such as mitigating biases in data to ensure fairness and reliability.

What's next for Deluminator

A robust misinformation detection system can include real-time news scraping to analyze trends using NLP, Explainable AI (XAI) for transparency via attention visualization, multi-language support with models like mBERT, and deepfake detection using AI tools like FaceForensics++. These features enhance accuracy, trust, and multimedia verification.