Machine learning (ML) has progressed rapidly during the past decade and the\nmajor factor that drives such development is the unprecedented large-scale\ndata. As data generation is a continuous process, this leads to ML model owners\nupdating their models frequently with newly-collected data in an online\nlearning scenario. In consequence, if an ML model is queried with the same set\nof data samples at two different points in time, it will provide different\nresults.\n In this paper, we investigate whether the change in the output of a black-box\nML model before and after being updated can leak information of the dataset\nused to perform the update, namely the updating set. This constitutes a new\nattack surface against black-box ML models and such information leakage may\ncompromise the intellectual property and data privacy of the ML model owner. We\npropose four attacks following an encoder-decoder formulation, which allows\ninferring diverse information of the updating set. Our new attacks are\nfacilitated by state-of-the-art deep learning techniques. In particular, we\npropose a hybrid generative model (CBM-GAN) that is based on generative\nadversarial networks (GANs) but includes a reconstructive loss that allows\nreconstructing accurate samples. Our experiments show that the proposed attacks\nachieve strong performance.\n