The detection of galactic binaries as sources of gravitational waves promises an unprecedented wealth of information about these systems, but also raises several challenges in signal processing. In particular the large number of expected sources and the risk of misidentification call for the development of robust methods. We describe here an original non-parametric reconstruction of the imprint of galactic binaries in measurements affected by instrumental noise typical of the space-based gravitational wave observatory LISA. We assess the impact of various approaches based on sparse signal modelling and focus on adaptive structured block sparsity. We carefully show that a sparse representation gives a reliable access to the physical content of the interferometric measurement. In particular we check the successful fast extraction of the gravitational wave signal on a simple yet realistic example involving verification galactic binaries recently proposed in LISA data challenges.