LDCDroid: Learning data drift characteristics for handling the model aging problem in Android malware detection

Abstract

The dynamic and evolving nature of malware applications can lead to deteriorating performance in malware detection models, a phenomenon known as the model aging problem. This issue compromises the model’s effectiveness in maintaining mobile security. Model retraining have proven effective in enhancing performance on previously unseen applications. However, the substantial need for annotated data remains a significant challenge in acquiring accurate ground truth for model retraining. Therefore, this paper introduces a new method to address the model aging problem in Android malware detection(AMD). To alleviate the burden of manual annotation, our approach incorporates pseudo-labeled data into the retraining process. Specifically, we introduce a novel method for evaluating the data drift scores of newly emerged samples by learning their data drift characteristics. These scores guide the usage of pseudo-labeled and true-labeled data for retraining the model. Our method significantly reduces the resources required for annotation while maintaining the efficacy of malware detection. In long-term datasets, we demonstrate the efficacy of our models through a series of experiments. Results indicate that our method enhances the F-score by approximately 26% in predicting unseen malware over a span of nine years.