In the Internet, becase of its huge scale and distributed administration, it is sometime difficult to measure internal states of the network directly, and thus, statistical/indirect ways to infer network-internal characteristics from end-to-end measurement have been studied. In particular, inferring link losses from end-to-end measurement along a tree-like path topology has been successfully developed in recent researches, which should be followed by deployment in practical uses. In this paper, therefore, we present an experiment of inferring link losses using end-to-end packet pair probing in the actual Internet, and show that it can infer link loss rates within 1% relative-errors in various network conditions. We also address applications for the inference method.