💻 computer science

F2IND-IT! -- Multimodal Fuzzy Fake Indian News Detection using Images and Text

This paper presents F2IND-IT, a novel multimodal framework that integrates ResNet-50 for image analysis, DistilBERT for text processing, and an Adaptive Neuro-Fuzzy Inference System (ANFIS) with attention-based fusion to achieve superior fake news detection performance on Indian media compared to existing methods.

Original authors: Kushal Trivedi, Murtuza Shaikh, Khushi Singh, Jeevaraj S.

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kushal Trivedi, Murtuza Shaikh, Khushi Singh, Jeevaraj S.

Original paper licensed under CC BY 4.0 (http://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 you are a detective trying to solve a mystery in a bustling, chaotic marketplace called India. In this marketplace, news travels faster than a runaway rickshaw. Sometimes, the news is true, but often, people are spreading stories that are half-true, exaggerated, or completely made up to stir up emotions or get clicks. This is the problem of "fake news."

The paper you shared introduces a new detective tool called F2IND-IT! (which stands for Fuzzy Fake Indian News Detection using Images and Text). Here is how this tool works, explained in simple terms:

1. The Problem: Too Many Stories, Too Many Lies

In India, almost everyone has a smartphone now. While this is great for sharing real news, it's also a playground for liars. They can edit photos, write emotional stories, or mix real facts with fake ones. Checking every single story by hand is impossible because there are too many, and it takes too long. The authors noticed that while computers are good at checking text or pictures, they struggle when they have to check both together specifically for Indian news.

2. The Solution: A Three-Part Detective Team

The F2IND-IT! model is like a detective team with three specialized members working together to decide if a news story is real or fake.

Member A: The Picture Expert (ResNet-50)

Think of this as a forensic artist who looks at the photos in the news.

What it does: It uses a powerful brain (a neural network called ResNet-50) to scan the image. It doesn't just see a "cat"; it looks for signs that the photo was edited, cropped, or taken out of context.
The result: It turns the picture into a long list of numbers (a "feature vector") that describes the visual truth.

Member B: The Story Reader (DistilBERT)

This member is a super-fast librarian who reads the text.

What it does: It uses a smart language model (DistilBERT) to read the headline and the article. It understands the tone, the words used, and the context. It knows the difference between a serious report and a clickbait headline designed to make you angry.
The result: It also turns the text into a long list of numbers representing the story's meaning.

Member C: The Wise Judge (ANFIS)

This is the most unique part of the team. Imagine a wise judge who doesn't just say "Yes" or "No," but understands shades of gray.

The "Fuzzy" Logic: In real life, things aren't always 100% black or white. A story might be 80% true but have one fake detail. This judge uses "Fuzzy Logic" (specifically a system called ANFIS) to handle these gray areas. It looks at the evidence from the Picture Expert and the Story Reader and calculates a "reliability score."
How it works: It uses a set of rules (like "If the picture looks edited AND the text is emotional, then it's likely fake") to make a final decision.

3. The Magic Glue: The Attention Mechanism

Sometimes, the Picture Expert might be confused (maybe the photo is blurry), but the Story Reader is very clear. Other times, the text is vague, but the photo is obviously fake.

The Solution: The model has a special "Attention" switch. It acts like a conductor in an orchestra, deciding how loudly to listen to the Picture Expert versus the Story Reader. If the photo is missing, it listens only to the text. If the text is weak, it weighs the photo more heavily. This ensures the final decision is balanced.

4. The Training Ground

To learn how to be a good detective, this team was trained on a massive library of Indian news called the IFND dataset.

The Data: They looked at over 56,000 news articles (though they had to throw some away because the photos were missing).
The Challenge: The data was tricky. There were way more real news stories than fake ones (about 24,000 real vs. 600 fake). The model had to learn to spot the rare fake ones without getting confused by the sheer volume of real ones.

5. The Results: A High-Scoring Detective

After training, the F2IND-IT! team was tested, and the results were impressive:

Accuracy: It got the right answer about 97.7% of the time.
Comparison: It did significantly better than older methods that only looked at text or only looked at pictures.
The "Fuzzy" Advantage: When the researchers tested the model without the wise judge (the ANFIS part), the score dropped. This proved that the fuzzy logic was essential for handling the tricky, gray-area cases.

Summary

In short, F2IND-IT! is a smart computer program designed specifically to catch fake news in India. It doesn't just read the words or look at the pictures; it combines both, uses a "fuzzy" brain to handle uncertainty, and weighs the evidence carefully. The result is a highly accurate system that can spot lies in a sea of information, helping to keep the news marketplace honest.

Technical Summary: F2IND-IT! Multimodal Fuzzy Fake Indian News Detection

Problem Statement
The paper addresses the critical challenge of detecting fake news within the Indian media landscape. While fake news detection is a global issue, the authors argue that existing approaches often fail to account for the specific nuances of Indian content, which is characterized by high diversity, regional variations, and significant political polarization. The World Economic Forum has identified India as having the highest global risk for misinformation. Current detection methods are often unimodal (relying solely on text or images) or lack robustness against the specific types of manipulation found in Indian news, such as image editing, clickbait, and emotionally charged language. Furthermore, there is a noted gap in credible, comprehensive datasets and automated tools specifically designed for Indian news articles.

Methodology
The authors propose F2IND-IT! (Fuzzy Fake Indian News Detection using Images and Text), a novel multimodal deep learning framework that integrates visual and textual features with a fuzzy inference system. The architecture consists of three primary stages:

Feature Extraction:
- Visual Modality: A pre-trained ResNet-50 Convolutional Neural Network (CNN) is employed to extract high-level visual features from news images. The final classification layer is removed, and the backbone is fine-tuned to output a 2048-dimensional feature vector via global average pooling.
- Textual Modality: DistilBERT is utilized as a text encoder to generate semantic embeddings. The input text is tokenized with [CLS] and [SEP] tokens. A mean pooling operation, weighted by the attention mask, converts the sequence of token embeddings into a fixed-size 768-dimensional sentence representation.
Attention-Based Fusion:
- To align the modalities, both the visual (2048-dim) and textual (768-dim) embeddings are projected into a common 512-dimensional space using Multi-Layer Perceptrons (MLPs).
- A lightweight, learnable attention mechanism calculates attention logits for each modality. These scores are normalized via softmax to assign dynamic weights, allowing the model to adaptively balance the contribution of text and images. The weighted embeddings are aggregated into a unified 512-dimensional representation.
Fuzzy Inference Classification (ANFIS):
- The fused features are fed into an Adaptive Neuro-Fuzzy Inference System (ANFIS) rather than a standard dense classification layer.
- Fuzzification: The input features are mapped using two learnable Gaussian membership functions.
- Rule Layer: The system evaluates $2^4 = 16$ fuzzy rules based on the product of membership values.
- Defuzzification: A weighted sum of the normalized firing strengths and rule outputs generates the final confidence score, passed through a sigmoid activation to produce a probability of the news being fake.

Key Contributions

Novel Architecture: The paper introduces the first known framework to combine multimodal deep learning (ResNet-50 + DistilBERT) with a neuro-fuzzy inference system (ANFIS) specifically for Indian fake news detection.
Dataset Utilization: The study leverages the IFND (Indian Fake News Dataset), a multimodal dataset containing 56,713 image-text pairs from 2013–2021. After preprocessing to handle missing images and resolution constraints, the final experimental set comprised 25,195 examples (heavily imbalanced with 24,576 real and 619 fake articles).
Handling Imbalance: The authors employ a custom loss function combining binary cross-entropy, Huber loss, and focal loss to address the severe class imbalance. They also utilize stratified 5-fold cross-validation and Macro-F1 scores to ensure fair evaluation across classes.

Results
The model was evaluated on the IFND dataset and compared against various baseline configurations (unimodal CNNs, LSTMs, and other multimodal setups).

Performance: F2IND-IT! achieved an Accuracy of 97.69%, a Macro-F1 score of 97.35%, a Precision of 97.89%, and a Recall of 98.01%.
Ablation Studies:
- Modality: Unimodal text models (Bi-LSTM) reached 92.70% accuracy, while unimodal image models (ResNet-50) reached only 76.60%. The multimodal approach significantly outperformed unimodal baselines.
- Encoder: DistilBERT (97.69%) performed comparably to the larger BERT model (97.71%) but with lower computational overhead.
- ANFIS Impact: Removing the ANFIS layer resulted in a drop in accuracy to 96.73%, demonstrating the value of the fuzzy inference layer in refining the classification boundary.

Significance and Claims
The authors claim that F2IND-IT! offers a robust solution for the specific challenges of Indian misinformation by effectively capturing complementary features from both text and images. The integration of fuzzy logic allows for the handling of uncertainty and non-linear trends inherent in fake news detection. The study highlights that while the model performs well, the reliance on expert-defined fuzzy partitions in the current ANFIS implementation is a limitation. The authors suggest that future work could focus on data-driven approaches to automatically identify fuzzy cluster centroids and spreads, thereby reducing the need for expert intervention in rule formation. The paper concludes that the proposed architecture consistently outperforms previous research on the IFND dataset across all evaluation metrics.