International Association
for Cryptologic Research

We found a statistical weakness in the Spritz algorithm designed by Ronald L. Rivest and Jacob C. N. Schuldt. For N=8: Prob(output(x)=output(x+2)) = 1/N + 0.000498. The bias becomes statistically significant (for N=8) after observing about 2^21.9 outputs. Analogous bias occurs for N=16. We propose an algorithm (VMPC-R) which for N=8 produced 2^46.8 (31 million times more) outputs which remained undistinguishable from random in the same battery of tests. Supported by a series of additional statistical tests and security analyses we present VMPC-R as an algorithm we hope can be considered a worthwhile replacement for RC4.