Skip to content
Crypto.world

Research

Post-Quantum Authentication of IoT Data on Ethereum: Signatures, the Calldata Floor, and Device-Signed Merkle Roots

Quantifies post-quantum signature costs for Ethereum IoT anchoring, showing the EIP-7623 calldata floor—not verification—dominates, and ships an LMS/HSS verifier.

Post-Quantum Authentication of IoT Data on Ethereum: Signatures, the Calldata Floor, and Device-Signed Merkle Roots
arXiv
2611.08108
Published
2026-11-05
Authors
Ming-Fong Sie

AI summary

Overview

Research area: Post-quantum cryptography for blockchain systems — specifically, the on-chain cost of authenticating IoT sensor data on Ethereum under current and scheduled gas rules.

Technical level: Advanced. The paper assumes familiarity with Ethereum gas accounting (EIP-7623, EIP-7825, EIP-7976, EIP-4844 blobs), NIST post-quantum signature families, and EVM-level measurement methodology.

Scope (one sentence): The paper measures, end to end on a real Osaka-rule EVM, what it costs to authenticate IoT records on Ethereum with post-quantum signatures — per record versus amortized through a device-signed Merkle root — and analyzes how the calldata floor, key placement and data availability determine that cost.

What This Paper Is About

Field IoT devices deployed today will still be producing records when quantum computers may be able to break the elliptic-curve signatures that currently authenticate them, and their firmware and key infrastructure are hard to update remotely. The paper asks whether a post-quantum authentication path can be added to an Ethereum anchoring contract now, without protocol changes, and what it costs under the current execution rules including the EIP-7623 calldata floor and the EIP-7825 transaction cap. Its central argument is that signing every record is priced by signature and key size rather than by verification, so a design in which each device signs one Merkle root over many records is what makes post-quantum authentication practical today.

Key Contributions

  1. Device-side cost measurements for ML-DSA, Falcon, SLH-DSA, LMS, HSS and XMSS from C benchmarks on x86-64 and ARM64 cores plus published Cortex-M4 measurements, reported together with uplink bytes and memory footprint per design.
  2. A new standard-compliant LMS/HSS verifier in Solidity, conforming to RFC 8554 and SP 800-208 for SHA-256 (n = 32) and SHA-256/192 (n = 24), validated against an independent implementation and executed for real, with its worst case derived from a regression on Winternitz chain steps; alongside reproduced costs of the ZKNOX ML-DSA and Falcon verifiers.
  3. End-to-end transaction costs on an Osaka-rule EVM for per-record signatures and four device-signed-root variants, with replay protection, benchmarked against ECDSA and P-256 in the same design, including hybrid ECDSA + post-quantum signatures and pricing under current rules, the scheduled Glamsterdam repricing, and proposed precompiles.
  4. A validated closed-form cost model (fitted per-device term A and per-leaf term L) with break-even formulas for choosing batch size k, plus an analysis of data-availability tiers and of alternatives to contract-level verification (STARK proof aggregation and optimistic anchoring).

Main Findings

  • The calldata floor, not verification, sets the price of per-record signatures. With the proposed ML-DSA-44 precompile, one record still costs 112,490 gas at 50 records per transaction — 8.7 times ECDSA's 12,971 gas — because every byte of signature and key costs 40 gas when the EIP-7623 floor binds, and verification adds nothing as long as execution stays below 6T. A Falcon-512 record costs 39,513 gas (3.0 times ECDSA).
  • Cheaper verification cannot fix per-record costs. Finding 1 of the paper states that cheaper precompiles alone cannot make per-record post-quantum authentication competitive with ECDSA; the decisive limit becomes capacity, about 530 ML-DSA-44 records per 60-million-gas block.
  • Device-signed roots are deployable today. With a standard Solidity LMS verifier and only the root posted (DR-0 at k = 100), a record costs about 3,350 gas (4,963 in the worst case), against 245 gas for ECDSA in the same design; with every leaf published (DR-L at k = 1,000) it costs 1,791 gas, only 22% above ECDSA (19% with n = 24).
  • The LMS verifier is comparatively cheap and hash-only. Its mean is 268,677 gas over 40 signatures (range 246,817–294,342), four times less than the optimized NIST ML-DSA verifier and sixteen times less than the Keccak-based XMSS verifier reported by Kysil et al.; cost is linear in Winternitz chain steps at 111,315 gas plus 322 gas per step (R² = 0.996), with 990 steps in the worst case costing 429,900 gas.
  • Batching removes the uplink penalty. An ML-DSA-44 device transmits 2,536 bytes per record with per-record signatures versus 140 bytes at k = 100 and 118 bytes at k = 1,000 with device-signed roots — an 18- to 22-fold reduction.
  • Scheduled repricing hurts floor-bound designs, not hash-based ones. Under Glamsterdam's EIP-7976 floor (64 gas per byte), floor-bound designs become 1.4 to 1.9 times more expensive while LMS is unaffected; LMS's premium over ECDSA falls from 3.0 to 2.0 times with leaves posted and from 13.7 to 10.1 times with roots only. A hypothetical 96/96 floor would raise floor-bound designs 2.0–2.8 times and, for the first time, LMS with posted leaves by 27%.
  • Precompiles and hybrid signatures are cheap at the margin. With proposed precompiles and the key stored on-chain (DR-0C at k = 100), ML-DSA-44 costs 1,031 gas per record and Falcon-512 387 gas; a hybrid ECDSA and ML-DSA signature adds only 2.5% in that design and at most 2% in DR-L.
  • Break-even batch sizes follow a simple rule. For the premium over ECDSA to fall below 25% with leaves posted, Falcon-512 with a precompile needs about 120 records per root (fitted 119), ML-DSA-44 about 390 (389), LMS about 800–900 (893 for n = 32), and NIST ML-DSA in Solidity about 3,700 (3,699).
  • On-device signing is an energy and uplink question, not a throughput one. On a Cortex-M4, ML-DSA-44 signs in 23–72 ms depending on stack budget, Falcon-512 in about 80 ms with 11 KB of RAM, SLH-DSA-128f in 2.2 s and SLH-DSA-128s in 46 s; per-record signing every ten minutes would be feasible even for SLH-DSA. Reference LMS without tree caching takes 22.6 s on the Cortex-M4 for height 10.
  • Stateful key budgets are ample under daily roots. An LMS tree of height 10 lasts 1,024 days (2.8 years), height 15 about 90 years, and a two-level HSS of two height-10 trees about 2,900 years; persisting the counter once per epoch means 365 flash writes per year instead of about 52,000 for a record every ten minutes.
  • Key storage pays off after six submissions. Storing an ML-DSA-44 key on-chain costs 309,403 gas more than registering a commitment (383,973 versus 74,570) and saves about 52,000 gas per root submission.
  • On rollups the scheme choice nearly disappears. ML-DSA-44 per-record signatures post 2,965 bytes per record against 28–73 bytes with device-signed roots (40- to 100-fold less); DR-0 with LMS posts 29.5 bytes and DR-0C with ML-DSA-44 posts 27.9 bytes, so where execution is cheap the choice of scheme barely matters.

Methodology in Plain English

The author builds a realistic anchoring contract — one that registers devices, stores a 32-byte key commitment, enforces monotonic per-device sequence numbers and epochs for replay protection, and has no administrator or upgrade key — and then measures real signed EIP-1559 transactions running on the EthereumJS VM with the Osaka hardfork rules active. Genuine keys and signatures go into calldata: ECDSA is verified for real with ecrecover, and LMS/HSS with the author's own verifiers. For schemes whose Solidity verifiers would blow the transaction cap, the contract instead calls an empty account with the full input, so everything except verification is still measured (including building the input in memory); the known call cost is then subtracted and the precompile or verifier price added. Verification prices come from reproducing open-source verifiers in Foundry using each repository's own settings and test vectors. Device timings come from liboqs C benchmarks on an x86-64 core and an ARM64 Neoverse N2 core, plus published Cortex-M4 figures. Because LMS cost depends on the message, forty signatures are regressed against chain-step count and the fit is evaluated at the observed maximum. Transaction gas is recomputed with the EIP-7623 formula, which matched the VM on all 2,021 stub transactions, and ten isolated LMS signatures predicted a ten-device transaction within 0.07%.

Why This Matters

Impact on research. The paper reframes post-quantum blockchain cost from a verification-gas problem to a data-cost problem: once the calldata floor binds, execution gas is effectively free and the binding constraint is signature and key bytes. It also contributes a standard-compliant (RFC 8554 / SP 800-208) LMS/HSS verifier, which matters because earlier hash-based EVM verifiers used Keccak instead of the standardized hash functions and therefore cannot verify signatures from standard signers or hardware.

Real-world applications:

  • Supply-chain finance: anchoring agricultural production records as tamper-evident evidence backing a loan, where a contract must release an instalment the moment a milestone is anchored, without relying on provers or challenge windows.
  • Agricultural monitoring camera fleets, which produce the 116-byte records the system model is built around.
  • Long-lived industrial and utility sensors whose firmware and key material cannot be updated remotely before quantum attacks become feasible.
  • Rollup-based IoT data markets, where the analysis shows that device-signed roots cut posted data by 40- to 100-fold and largely erase the choice of signature scheme.

Industry relevance. Deployers must decide now whether to be crypto-agile, and the paper gives concrete numbers: cost per record in ETH at 0.1, 1 and 10 gwei base fee (ML-DSA-44 per record with precompile: 2.3×10⁻⁵, 1.2×10⁻⁴ and 1.1×10⁻³ ETH; LMS with k = 100: 6.7×10⁻⁷, 3.7×10⁻⁶ and 3.4×10⁻⁵ ETH), block capacity (about 4,600 per-record ECDSA records, 530 per-record ML-DSA-44, and about 58,000 device-signed ML-DSA-44 roots at k = 100), and explicit warnings that stateful schemes require monotonic counters in hardware and must never be cloned or rolled back.

Future Directions

  • Blob-based root posting is estimated, not measured. DR-B is priced at the EIP-7918 reserve (about 2 gas per leaf for full blobs, 8 for a 1,000-leaf batch, 82 for a 100-leaf batch, multiplied by the market-to-reserve ratio), and the author states explicitly that blob transactions were not executed.
  • The Glamsterdam transition needs re-measurement. EIP-2780 (resource-based intrinsic gas), EIP-8037 (separately metered state creation, outside the transaction cap) and EIP-8038 (higher state-access costs) will change registration, key storage and the per-batch storage write, shifting the six-submission break-even for on-chain key storage.
  • Choosing between contract-level verification and aggregation. STARK proof aggregation breaks even at roughly 60–90 devices against DR-0C ML-DSA-44 (about 103,000 gas per device root) and 18–27 devices against LMS (about 335,000 gas), leaving the trade-off between on-chain gas and prover hardware, latency and trust assumptions open.
  • Fleet-scale and application-level validation. The paper notes that whether a per-record cost of 10⁻⁴ ETH is acceptable depends on the application, and it leaves the latency of batching and the data-availability properties of each variant to be matched against specific deployments.

Target Audience

Blockchain protocol and EIP reviewers assessing calldata pricing and post-quantum precompile proposals; embedded and IoT security engineers choosing signature schemes, hardware counters and firmware update paths for long-lived devices; smart-contract architects designing data-anchoring, supply-chain or evidence systems; and applied cryptography researchers interested in the practical on-chain cost of NIST post-quantum standards.

Authors’ abstract

Field devices deployed today will still be producing records when quantum computers may break their signatures, so systems that anchor IoT data on blockchains need a post-quantum (PQ) authentication path now. We measure its cost on Ethereum under the current execution rules, including the EIP-7623 calldata floor and the EIP-7825 transaction cap, and under the calldata repricing scheduled for the Glamsterdam upgrade. We combine device-side measurements of ML-DSA, Falcon,…

The full abstract on arXiv
Read the original paper