Popular web application features, like autosuggestion and auto-correction, are well-known to amply reveal input actions entered by their users to any eavesdroppers, by generating sufficiently unique network activity bursts on actions. Some techniques, like k-anonymity and l-diversity, already exist in the literature for achieving different levels of privacy of user actions by making the generated network bursts for different actions indistinguishable from each other. We present a method that blends network bursts of not entire actions, but their probabilistic portions, with each other. Our method thus results in the maximum possible amount of l-diversity, namely k-diversity, and is a sianificant improvement over the existing ones.