A Hybrid Temporal Lambda Layer Embedded in Autoencoder Neural Network for Fake News Detection
Many social media applications use different animated or morphed images to make fake news viral. Recognition of text from images for their classification as real or fake requires a neural network. BERT (Bidirectional Encoder Representation Transformer) or MLP-based (Multi-Layer Perceptron) algorithms are successful when working with textual data alone. However, the system needs to extract the sequential text from the images to identify the semantic meaning of the content before the classification process. The dataset utilized was acquired from The Indian Fake News Dataset (IFND) contains text and visual data from 2013 to 2021. The data includes both visual and textual information, as well as 126k data points obtained from millions of users. In the proposed model, a squeezed lambda is implemented to process the data in the three forms of verbal tenses, i.e., past to future and future to past. In the lambda layer, temporal classification is performed by applying two bidirectional LSTM (Long Short Term Memory) layers based on the retuning sequences of the character list available in the dataset. It also computes the batch cost of every iteration and reduces them based on the ratio of prediction and input class labels available. To ensure that the suggested technique is more accurate than the current approach, a validation was undertaken, resulting in a +0.5 increase in accuracy over the BERT (Bidirectional Encoder Representation Transformer) model. Hence, the proposed method has achieved higher accuracy than existing algorithms. Than existing algorithms.
Volume & Issue
Vol. Volume 15 / Iss. Issue 1