Widespread use and general purpose computing capabilities of next generation smartphones make them the next big targets of malicious software (malware) and security attacks. Given the battery, computing power, and bandwidth limitations inherent to such mobile devices, detection of malware on them is a research challenge that requires a different approach than the ones used for desktop/laptop computing. We present a novel probabilistic diffusion scheme for detecting anomalies possibly indicating malware which is based on device usage patterns. The relationship between samples of normal behavior and their features are modeled through a bipartite graph which constitutes the basis for the stochastic diffusion process. Subsequently, we establish an indirect similarity measure among sample points. The diffusion kernel derived over the feature space together with the Kullback-Leibler divergence over the sample space provide an anomaly detection algorithm. We demonstrate its applicability in two settings using real world mobile phone data. Initial experiments indicate that the diffusion algorithm outperforms others even under limited training data availability.