Publish/subscribe communication is a common architectural design pattern in component-based software systems used in many of today's cyber-physical systems to exchange information between distributed software components. These systems typically deal with an increased number of data transfers, with a risk of lacking resources. Our recent domain analysis for a lane-following algorithm of a self-driving miniature car unveiled that the actual information increment between two subsequently sent packets is often small. Such scenario enables possibilities for a more efficient data exchange by avoiding redundant and/or unnecessary information transfer. In this paper, we propose and evaluate our concept for self-adaptive data marshalling that transparently adapts data types in messages to be exchanged by analyzing the actual information increment. The approach could reduce the bandwidth usage by more than 50% in comparison to the current approach, and by approximately 33% compared to the use of the general-purpose compression library zlib.