1. Underlying statistic is broken 2. Should add noise using noisy histograms on each level rather than adding Laplace noise to weights of nodes individually.