A Multi-Lingual Cyberbullying Detection and Blockchain-Integrated Privacy-Preserving Federated Learning Framework
This paper proposes a privacy-preserving, blockchain-integrated federated learning framework that utilizes NLP and deep learning models, particularly BiLSTM, to detect multilingual cyberbullying in English, Bangla, and Banglish while ensuring data security and user privacy.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine the internet as a giant, bustling digital playground where billions of people chat, share memes, and tell stories every second. It's a place of connection, but like any playground, it has its bullies. These aren't just kids shoving others on the swings; they are online harassers using hurtful words in dozens of different languages to make people feel small, scared, or angry. For years, scientists have tried to build "digital security guards" (algorithms) to spot this bad behavior. Usually, these guards work by gathering all the chat logs into one giant, central vault to study them. But here's the catch: putting everyone's private conversations into one big vault is risky. If the vault gets hacked, or if someone sneaks a peek, everyone's secrets are exposed. To fix this, researchers are now exploring two new tools: Federated Learning, which is like letting students study for a test in their own homes and only sharing their answers (not their notes) with the teacher, and Blockchain, which acts like an unbreakable, public ledger that records every step of the process so no one can alter or change the history. The big question is: Can we build a super-smart bully detector that learns from many different languages without ever seeing the private messages, while keeping a secure, tamper-proof record of how it learned?
This paper introduces a clever new system that tries to do exactly that. The authors, a team from the Bangladesh University of Business and Technology, built a framework designed to catch cyberbullying in three specific types of text: English, Bangla (the language of Bangladesh), and "Banglish" (a fun mix of both). Instead of stealing all the data to a central server, they used Federated Learning. Imagine a classroom where every student has their own private notebook of social media posts. The teacher (the central server) sends out a "smart brain" (a computer model) to every student. Each student trains this brain on their own private notebook, learning to spot mean words, but they never show their notebook to the teacher or the other students. They only send back the "lessons learned" (updated math weights) to the teacher. The teacher then combines all these lessons to create a super-smart global brain that knows how to spot bullying in all three languages, all while keeping every student's private data safe in their own pockets.
To make sure no one tries to alter or swap out the lessons during this process, the team added Blockchain technology. Think of this as a magical, unerasable diary where every time a student sends in their lessons, a new page is added. This page is locked with a special digital seal (SHA-256 hashing) and verified by a puzzle-solving game (Proof-of-Work). If someone tries to sneak in a fake lesson or change a past entry, the whole chain breaks, and the system knows something is wrong. This ensures that the training process is transparent and secure.
The researchers tested four different "brains" to see which one was best at this job: BiLSTM, LSTM, Random Forest, and XGBoost. They ran experiments using 21,204 social media posts collected between 2023 and 2025. The results were clear: the BiLSTM model was the champion. It achieved a stunning accuracy of 98.43% in the final test, meaning it correctly identified bullying almost every single time. The LSTM model came in second with 94.43% accuracy, followed by XGBoost at 93.20%, and Random Forest lagged behind at 81.76%. The paper suggests that the BiLSTM model won because it can read text in two directions at once—looking at the words before and after a specific phrase to understand the full context, which is crucial for catching tricky, mixed-language insults.
While the system worked incredibly well in these tests, the authors are careful to note that it isn't a magic wand that solves everything instantly. They point out that this high level of security and intelligence comes with a cost: it requires a lot of computing power and communication between the devices. The paper suggests that while this framework is a robust and secure solution for the future, researchers might need to find lighter, faster versions of these models to make them run smoothly on everyday devices without slowing down the internet. Ultimately, the study proves that we can build a bully detector that is both incredibly smart and fiercely protective of user privacy, without needing to sacrifice one for the other.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.