Analysis of random noise generated by graphic processing units

Abstract

Random number generators are essential in modern cryptography. The security of a cryptographic scheme can be achieved under the assumption that the system uses ideal random numbers to produce sensitive security parameters such as encryption keys and initial vectors. The weakness of the random number generator makes the entire cryptographic system insecure. In particular, the lack of entropy sources leads to predictable output random bits so that secret information can be guessed by malicious attackers. Therefore, it is important to collect sufficient entropy from physical noise sources. In this paper, we consider graphics processing units (GPUs) as an entropy source. From the race conditions in the parallel computations on a GPU, we can harvest sufficient entropy for cryptography. Using the entropy estimations in NIST SP 800-90B, the amount of entropy is estimated and compared with other physical sources.