Watermarking has been used on multimedia for a variety of applications, among which intellectual property protection stands out. However, and in spite of a growing need, very few attempts have been made at using watermarking to protect the intellectual property of al- phanumerical databases. We present in this paper a watermarking sys- tem for numerical databases, which is the first one of its kind to pre- serve means and variances of attributes in the database. Given a wa- termarked dataset and a candidate watermark, the recovery algorithm makes a decision whether the candidate watermark is embedded or not in the dataset. The probabilities of false positive and false negative can be made arbitrarily small by proper choice of a security parameter. We give an analytical expression for the information or data utility loss caused by watermark embedding. The proposed system can be made arbitrarily robust against noise addition. We also give empirical results showing a noise addition attack trying to remove the watermark must cause a dis- tortion (and thus a loss of data utility) much larger than the watermark itself.