Improved torsion point attacks on SIDH variants

TL;DR

This paper enhances torsion point attacks on SIDH, significantly lowering the difficulty of certain cryptographic problems, and proposes SIDH variants resistant to these improved attacks, without compromising SIKE security.

Contribution

It strengthens torsion point attack techniques using dual and Frobenius isogenies and constructs SIDH variants designed to resist these attacks.

Findings

Complete break of n-party group key exchange for 6+ parties

Quantum attack for 3+ parties with improved complexity

Implementation of attacks in Magma for 6 parties

Abstract

SIDH is a post-quantum key exchange algorithm based on the presumed difficulty of finding isogenies between supersingular elliptic curves. However, SIDH and related cryptosystems also reveal additional information: the restriction of a secret isogeny to a subgroup of the curve (torsion point information). Petit (2017) was the first to demonstrate that torsion point information could noticeably lower the difficulty of finding secret isogenies. In particular, Petit showed that "overstretched" parameterizations of SIDH could be broken in polynomial time. However, this did not impact the security of any cryptosystems proposed in the literature. The contribution of this paper is twofold: First, we strengthen the techniques of Petit by exploiting additional information coming from a dual and a Frobenius isogeny. This extends the impact of torsion point attacks considerably. In particular, our techniques yield a classical attack that completely breaks the n-party group key exchange of Azarderakhsh et al. for 6 parties or more, and a quantum attack for 3 parties or more that improves on the best known asymptotic complexity. We also provide a Magma implementation of our attack for 6 parties. We give the full range of parameters for which our attacks apply. Second, we construct SIDH variants designed to be weak against our attacks; this includes backdoor choices of starting curve, as well as backdoor choices of base-field prime. We stress that our results do not degrade the security of, or reveal any weakness in, the NIST submission SIKE.