TCMB: Cross-Model Multi-Level Cross-Attention Network with Taylor-Based Loss for multimodal Fake News Detection

Abstract

Detection of fake news leverages algorithms to classify and flag content that deviates from factual reporting. It's a critical task in the modern digital era, where incorrect information spreads easily across social media, websites, and messaging platforms. Though several methods exist for fake news detection, they suffer from limitations, like bias in training data, and high false positives and negatives. Therefore, a Multimodal Cross Attention Network with Taylor-based Cross Entropy Mean Bias (MMCN_TCMB) model is introduced in this paper for detecting multimodal fake news. At first, multimodal input, such as text and images, is acquired from the datasets. The textual content in the news posts is converted into tokens by Bidirectional Encoder Representations from Transformers (BERT). Then, the feature extraction process utilizes Word2Vec and Term Frequency-Inverse Gravity Moment (TF-IGM). Simultaneously, the image in the news post is subjected to Contrast Limited Adaptive Histogram Equalization-Histogram Equalization (CLAHE-HE), followed by the extraction of the ResNet feature. Subsequently, the resulting features are combined and fed into the MMCN with TCMB for detecting fake news. Here, the learning rule of MMCN is updated utilizing the TCMB loss function. The results reveal that the MMCN_TCMB obtains a better recall, precision, F1-score, and accuracy of 97.988%, 96.223%, 97.098%, and 97.436%. The code is available on: banbhrani84/TCMB-Cross_Model_Multi_Level_Cross_Attention_Network: Multimodal Cross Attention Network with Taylor-based Cross Entropy Mean Bias (MMCN_TCMB)