International Association
for Cryptologic Research

Masking is the main countermeasure against side-channel attacks on embedded devices. For cryptographic algorithms that combine Boolean and arithmetic masking, one must therefore convert between the two types of masking, without leaking additional information to the attacker. In this paper we describe a new high-order conversion algorithm between Boolean and arithmetic masking, based on table recomputation, and provably secure in the ISW probing model. We show that our technique is particularly efficient for masking structured LWE encryption schemes such as Kyber and Saber. In particular, for Kyber IND-CPA decryption, we obtain an order of magnitude improvement compared to existing techniques.

High-order masking countermeasures against side-channel attacks usually require plenty of randomness during their execution. For security against t probes, the classical ISW countermeasure requires O(t^2 s) random bits, where s is the circuit size. However running a True Random Number Generator (TRNG) can be costly in practice and become a bottleneck on embedded devices. In [IKL+13] the authors introduced the notion of robust pseudo-random number generator (PRG), which must remain secure even against an adversary who can probe at most t wires. They showed that when embedding a robust PRG within a private circuit, the number of random bits can be reduced to O(t^4), that is independent of the circuit size s (up to a logarithmic factor). Using bipartite expander graphs, this can be further reduced to O(t^(3+eps)); however the resulting construction is unpractical. In this paper we describe a practical construction where the number of random bits is only O(t^2) for security against t probes, without expander graphs; moreover the running time of each pseudo-random generation goes down from O(t^4) to O(t). Our technique consists in using multiple independent PRGs instead of a single one. We show that for ISW circuits, the robustness property of the PRG is not required anymore, which leads to simple and efficient constructions. For example, for AES we only need 48 bytes of randomness to get second-order security (t=2), instead of 2880 in the original Rivain-Prouff countermeasure; when implemented on an ARM-based embedded device with a relatively slow TRNG, we obtain a 50% speed-up compared to Rivain-Prouff.

Masking is an effective countermeasure against side-channel attacks. In this paper, we improve the efficiency of the high-order masking of look-up tables countermeasure introduced at Eurocrypt 2014, based on a combination of three techniques, and still with a proof of security in the Ishai-Sahai-Wagner (ISW) probing model. The first technique consists in proving security under the stronger t-SNI definition, which enables to use n = t+1 shares instead of n = 2t+1 against t-th order attacks. The second technique consists in progressively incrementing the number of shares within the countermeasure, from a single share to n, thereby reducing the complexity of the countermeasure. The third technique consists in adapting the common shares approach introduced by Coron et al. at CHES 2016, so that half of a randomized look-up table can be pre-computed for multiple SBoxes. We show that our techniques perform well in practice. In theory, the combination of the three techniques should lead to a factor 10.7 improvement in efficiency, for a large number of shares. For a practical implementation with a reasonable number of shares, we get a 4.8 speed-up factor for AES.

Masking is a very common countermeasure against side channel attacks. When combining Boolean and arithmetic masking, one must be able to convert between the two types of masking, and the conversion algorithm itself must be secure against side-channel attacks. An efficient high-order Boolean to arithmetic conversion scheme was recently described at CHES 2017, with complexity independent of the register size. In this paper we describe a simplified variant with fewer mask refreshing, and still with a proof of security in the ISW probing model. In practical implementations, our variant is roughly 25% faster.