Semantic-based Federated Defense for Distributed Malicious Attacks

Consumer electronics, particularly Android, has become the leading mobile ecosystem due to its accessibility and adaptability. However, the constant connectivity of Android apps makes them a target for malicious network attacks, leading to the potential theft of sensitive data and disruptions across various sectors, including the economy and healthcare. Developing a reliable, distributed malware detection system using training data from multiple sources is challenging. This is mostly due to privacy issues and a lack of consistent data. This paper proposes a semantic-driven Federated Learning (FL) approach using transformer-based transfer learning to defend against distributed malicious attacks. The semantic features of malicious scripts are examined using the Bidirectional Encoder Representations from the Transformers (BERT) model. Following that, Deep Neural Network (DNN) uses these semantic features for local training, resulting in local model updates for each client. After merging the local model updates from each client, the global server generates global weights and sends them to distant clients. The proposed approach is evaluated on two standard datasets, including CIC-AndMal2017 and CICMalDroid2020, and it obtains high detection accuracy of 99.38% and 99.14%, respectively. These findings encourage cybersecurity organizations to collaborate and develop a powerful distributed security system using private data. IEEE