TIF: Learning Temporal Invariance in Android Malware Detectors

The paper introduces TIF, a novel temporal invariant training framework designed to improve the stability of representations in Android malware detectors facing distribution drift. TIF utilizes multi-proxy contrastive learning and invariant gradient alignment to effectively manage temporal drift by organizing environments based on application observation dates. Experimental results demonstrate that TIF significantly enhances detection performance, especially during early deployment phases, outperforming existing state-of-the-art methods and addressing critical challenges in malware detection.