A Comparison of Post-Processing Techniques for Biased Random Number Generators

In this paper, we study and compare two popular methods for postprocessing random number generators: linear and Von Neumann compression. We show that linear compression can achieve much better throughput than Von Neumann compression, while achieving practically good level of security. We also introduce a concept known as the adversary bias which measures how accurately an adversary can guess the output of a random number generator, e.g. through a trapdoor or a bad RNG design. Then we prove that linear compression performs much better than Von Neumann compression when correcting adversary bias. Finally, we discuss on good ways to implement this linear compression in hardware and give a field-programmable gate array (FPGA) implementation to provide resource utilization estimates.