CFSL-BC: Compression-enabled federated split learning with blockchain for robust android malware detection

초록

Android applications have increased the difficulty of identifying advanced and obfuscated malware due to the rapid proliferation. Available federated and blockchain-based models are typically characterized by high communication overhead, low robustness, and insufficient feature diversity. In the study, Compression-Enabled Federated Split Learning with Blockchain (CFSL-BC) is proposed as a new framework that combines compression, split learning (SL), and blockchain to provide an effective, privacy-protected, and robust blockchain detection. The system also relies on built-in compression to reduce transportation costs, SL during joint model training, and smart contracts, with IPFS providing security features. In addition, Byzantine-robust aggregation and differential privacy defend against both poisoning and leakage attacks of the model. The results of experiments on the CIC-MalDroid 2020, Drebin, and Maloid-DS datasets show that CFSL-BC achieves 98.12% accuracy and incurs over 60% less communication overhead than the traditional federated framework. These findings validate CFSL-BC as a feasible and safe mechanism for detecting next-generation malicious Android software. Future research will focus on adaptive compression techniques, cross-domain applications in IoT systems, and a blockchain-based incentive scheme to further enhance the scalability, energy efficiency, and trustworthiness of participating nodes.

키워드