Paper 2018/1207

On the Decoding Failure Rate of QC-MDPC Bit-Flipping Decoders

Nicolas Sendrier and Valentin Vasseur

Abstract

Quasi-cyclic moderate density parity check codes allow the design of McEliece-like public-key encryption schemes with compact keys and a security that provably reduces to hard decoding problems for quasi-cyclic codes. In particular, QC-MDPC are among the most promising code-based key encapsulation mechanisms (KEM) that are proposed to the NIST call for standardization of quantum safe cryptography (two proposals, BIKE and QC-MDPC KEM). The first generation of decoding algorithms suffers from a small, but not negligible, decoding failure rate (DFR in the order of $10^{-7}$ to $10^{-10}$). This allows a key recovery attack presented by Guo, Johansson, and Stankovski (GJS attack) at Asiacrypt 2016 which exploits a small correlation between the faulty message patterns and the secret key of the scheme, and limits the usage of the scheme to KEMs using ephemeral public keys. It does not impact the interactive establishment of secure communications (e.g. TLS), but the use of static public keys for asynchronous applications (e.g. email) is rendered dangerous. Understanding and improving the decoding of QCMDPC is thus of interest for cryptographic applications. In particular, finding parameters for which the failure rate is provably negligible (typically as low as $2^{-64}$ or $2^{-128}$) would allow static keys and increase the applicability of the mentioned cryptosystems. We study here a simple variant of bit-flipping decoding, which we call step-by-step decoding. It has a higher DFR but its evolution can be modelled by a Markov chain, within the theoretical framework of Julia Chaulet's PhD thesis. We study two other, more efficient, decoders. One is the textbook algorithm. The other is (close to) the BIKE decoder. For all those algorithms we provide simulation results, and, assuming an evolution similar to the step-by-step decoder, we extrapolate the value of the DFR as a function of the block length. This will give an indication of how much the code parameters must be increased to ensure resistance to the GJS attack.