Paper 2024/276

Reduce and Prange: Revisiting Prange's Information Set Decoding for LPN and RSD

Abstract

The learning parity with noise (LPN) problem has been widely utilized in classical cryptography to construct cryptographic primitives. Various variants of LPN have been proposed, including LPN over large fields and LPN with regular noise, depending on the underlying space and the noise regularity. These LPN variants have proven to be useful in constructing cryptographic primitives. We propose an improvement to the Gaussian elimination attack, which is also known as Prange's information set decoding algorithm, for solving the LPN problem. Contrary to prevailing knowledge, we find that the Gaussian elimination attack is highly competitive and currently the best method for solving LPN over large fields. Our improvement involves applying partial Gaussian elimination repeatedly, rather than the whole Gaussian algorithm, which we have named the ``Reduce and Prange's algorithm". Moreover, we provide two applications of Reduce and Prange algorithms: One is the hybrid algorithm of ours and Berstein, Lange and Peters's algorithm at PQCrypto'08, and the other one is Reduce and Prange algorithm for LPN with regular noise. Last, we provide a concrete estimation of the bit-security of LPN variants using our Reduce and Prange's frameworks. Our results show that the bit-security of LPN over $\mathbb{F}_q$ is reduced by 5-11 bits when $\log q = 128$ compared to previous analysis by Liu et al. (will appear at Eurocrypt'24). Furthermore, we show that our algorithm outperforms recent work by Briaud and Øygard (Eurocrypt'23) and Liu et al. for certain parameters. It reduces the bit-security of LPN with regular noise by 5-28 bits.