Paper 2024/1081

Practical Non-interactive Multi-signatures, and a Multi-to-Aggregate Signatures Compiler

Abstract

In a fully non-interactive multi-signature, resp. aggregate-signature scheme (fNIM, resp. fNIA), signatures issued by many signers on the same message, resp. on different messages, can be succinctly ``combined'', resp. ``aggregated''. fNIMs are used in the Ethereum consensus protocol, to produce the certificates of validity of blocks which are to be verified by billions of clients. fNIAs are used in some PBFT-like consensus protocols, such as the production version of Diem by Aptos, to replace the forwarding of many signatures by a new leader. In this work we address three complexity bottlenecks. (i) fNIAs are costlier than fNIMs, e.g., we observe that verification time of a 3000-wise aggregate signature of BGLS (Eurocrypt'03), takes 300x longer verification time than verification of a 3000-wise pairing-based multisignature. (ii) fNIMs impose that each verifier processes the setup published by the group of potential signers. This processing consists either in verifying proofs of possession (PoPs), such as in Pixel (Usenix'20) and in the IETF'22 draft inherited from Ristenpart-Yilek (Eurocrypt'07), which costs a product of pairings over all published keys. Or, it consists in re-randomizing the keys, such as in SMSKR (FC'24). (iii) Existing proven security bounds on efficient fNIMs do not give any guarantee in practical curves with 256bits-large groups, such as BLS12-381 (used in Ethereum) or BLS12-377 (used in Zexe). Thus, computing in much larger curves is required to have provable guarantees. Our first contribution is a new fNIM called $$, it addresses both (ii) and (iii). It is as simple as adding Schnorr PoPs to the schoolbook pairing-based fNIM of Boldyreva (PKC'03). (ii) For a group of 1000 signers, verification of these PoPs is: $$ times faster than for the previous pairing-based PoPs; and $$ times faster than the Verifier's processing of the setup in SMSKR (and contrary to the latter, needs not be re-started when a new member joins the group). (iii) We prove a tight reduction to the discrete logarithm (DL), in the algebraic group model (AGM). Given the current estimation of roughly 128 bits of security for the DL in both the curves BLS12-381 and BLS12-377, we deduce a probability of forgery of $$ no higher than about $$ for a time $$ adversary. This reduction is our main technical contribution. The only related proof before was for an interactive Schnorr-based multi-signature scheme, using Schnorr PoPs. Our approach easily fills a gap in this proof, since we take into account that the adversary has access to a signing oracle even before publishing its PoPs. But in our context of pairing-based multi-signatures, extraction of the keys of the adversary is significantly more complicated, since the signing oracle produces a correlated random string. We finally provide another application of $$, which is that it can be plugged in recent threshold signatures without setup (presented by Das et al at CCS'23, and Garg et al at SP'24), since these schemes implicitly build on any arbitrary BLS-based fNIM. Our second contribution addresses (i), it is a very simple compiler: $$ (multi-to-aggregate). It turns any fNIM into an fNIA, suitable for aggregation of signatures on messages with a prefix in common, with the restriction that a signer must not sign twice using the same prefix. The resulting fNIA is post-quantum secure as soon as the fNIM is, such as Chipmunk (CCS'23). We demonstrate the relevance for Diem by applying $$ to $$: the resulting fNIA enables to verify 39x faster an aggregate of 129 signatures, over messages with $$ bits-long variable parts, than BGLS.

Note: See changelog (appendix E) for a list of what's changed. Expands and proves some results introduced in the 2023 version of 2020/1480.