The integration of mobile devices into adolescents’ daily lives is significant, making it imperative to prioritize their safety and security. With the imminent arrival of fast internet (6G), offering increased bandwidth and reduced latency compared to its predecessor (5G), real-time streaming of high-quality video and audio to mobile devices will become feasible. To effectively leverage the fast internet, accurately classifying Mobile Applications (M-APPs) is crucial to shield adolescents from inappropriate content, including violent videos, pornography, hate speech, and cyberbullying. This work introduces an innovative approach utilizing Deep Learning techniques, specifically Attentional Convolutional Neural Networks (A-CNNs), for classifying M-APPs. The goal is to secure adolescent mobile usage by predicting the potential negative impact of M-APPs on adolescents. The proposed methodology employs multiple Machine and Deep Learning (M/DL) models, but A-CNNs based on Bidirectional Encoder Representations from Transformers embeddings outperformed other models, achieving an average accuracy of 88.74% and improving the recall from 99.33% to 99.65%.

The integration of mobile devices into adolescents’ daily lives is significant, making it imperative to prioritize their safety and security. With the imminent arrival of fast internet (6G), offering increased bandwidth and reduced latency compared to its predecessor (5G), real-time streaming of high-quality video and audio to mobile devices will become feasible. To effectively leverage the fast internet, accurately classifying Mobile Applications (M-APPs) is crucial to shield adolescents from inappropriate content, including violent videos, pornography, hate speech, and cyberbullying. This work introduces an innovative approach utilizing Deep Learning techniques, specifically Attentional Convolutional Neural Networks (A-CNNs), for classifying M-APPs. The goal is to secure adolescent mobile usage by predicting the potential negative impact of M-APPs on adolescents. The proposed methodology employs multiple Machine and Deep Learning (M/DL) models, but A-CNNs based on Bidirectional Encoder Representations from Transformers embeddings outperformed other models, achieving an average accuracy of 88.74% and improving the recall from 99.33% to 99.65%.