Federated Defense for Malware Detection and Resilience Against Adversarial Attacks

Consumer electronics are substantially compromised by malware, which can traverse numerous operating systems and file formats. Considerable effort has been devoted to developing malware detection systems that employ machine learning and deep learning. However, these models are susceptible to adversarial attacks, where maliciously crafted inputs can bypass detection mechanisms. In this article, we present Fed-Adversarial, a novel technique for malware detection against adversarial attacks that employ intermittent clients-based federated learning. This method can improve adversarial attack detection while preserving data privacy for each client. The raw malware images are first normalized and converted to color to extract features efficiently. In addition, a wide range of adversarial examples is generated using normalized images to maximize evasion opportunities and reduce perturbations. Following this, adversarial examples are employed by deep convolutional neural networks during local training, resulting in local model updates (LMUs). After combining these LMUs, the global server produces global model updates, delivered to distant clients. The proposed approach is evaluated on standard datasets, including dumpware10, malimg, and MaleVis, and it obtains high detection accuracy of 99.18%, 98.12%, and 98.38%, respectively.