Mikael Carmona
Univ. Grenoble Alpes, CEA, LETI MINATEC Campus, F-38054 Grenoble, France

Viktor Fischer
Hubert Curien Laboratory, Jean Monnet University, Saint-Etienne, France; Faculty of Information Technologies, Czech Technical University, Prague, Czech Republic

Florian Pebay-Peyroula
Univ. Grenoble Alpes, CEA, LETI MINATEC Campus, F-38054 Grenoble, France

Romain Wacquez
Univ. Grenoble Alpes, CEA, LETI MINATEC Campus, F-38054 Grenoble, France; CEA-Leti, Mines Saint-Étienne, Equipe Commune, F-13541 Gardanne, France

Keywords: Random number generation, Parameterized stochastic models, Dedicated statistical tests, Physical isolation, Hardware monitoring

Abstract

Ring Oscillators (RO) are often used in true random number generators (TRNG). Their jittered clock signal, used as randomness source, originates from thermal and flicker noises. While thermal noise jitter is generally used as the main source of randomness, flicker noise jitter is not due to its autocorrelation. This work aims at qualitatively settling the issue of the influence of flicker noise in TRNGs, as its impact increases in newer technology nodes. For this, we built a RO behavioural model, which generates time series equivalent to a jittered RO signal. It is then used to generate the output of an elementary RO-TRNG. Despite general expectations, the autocorrelation inside the output bit stream is reduced when the amplitude of flicker noise increases. The model shows that this effect is caused by the sampling of the jittered signal by the second oscillator, which hides the behaviour of the absolute jitter, causes resetting of the perceived phase, and suppresses any memory effect. The inclusion of flicker noise as a legitimate noise source can increase the TRNG output bit rate by a factor of four for the same output entropy rate. This observation opens new perspectives towards more efficient stochastic models of the RO-TRNGs.