Deployment of Deep Learning Models to Mobile Devices for Spam Classification

Abstract

The advent of deep learning brings the possibility of better and faster applications in real world. In this work, deep learning models are used for application of spam classification in mobile devices. A Binary Classification model is trained with deep learning and is transformed to a graph using tensorflow and then, is converted to a protobuf file to be deployed on mobile devices. Instead of looking into the spam messages in an algorithmic way i.e. just with keywords, binary model deals with experience of learning and predicts if a text message is spam. The training was performed multiple times on resource-deficient devices and hyper-parameter optimization was performed to enhance the training accuracy to 99.87 %. The test accuracy of mobile application is 98.7 % and testing happens in real-time without any internet access. Our simulation shows that a model with an embedding layer (size 128), an LSTM layer (size 64, dropout 0.2) and a dense layer (sigmoid) yields the highest performance. Also, the comparative evaluation with state-of-the-art methods displayed that our model achieves higher accuracy.