Customer data privacy is known to be a factor which makes just-in-time data sharing and mining among enterprises challenging. Learning-from-abstraction is a recently proposed paradigm for privacy preserving distributed data mining where distributed local data sources are protected by probabilistic data abstraction. In this paper, we investigate the use of a normalized negative log likelihood together with the paradigm for quantifying the level of privacy protection, and studied theoretically the change of the privacy levels of the local data abstractions after being aggregated for global data analysis. Experiments on distributed data clustering with a synthetic data set were conducted on a service-oriented BPEL platform. The promising results obtained demonstrates the effectiveness of the adopted privacy measure.