|
PQ-WB-KEM: White-Box ML-KEM-768 1.0
PQ-WB-KEM: Toward a White-Box Construction of ML-KEM-768 with Arithmetic Masking for M2M Communications - Mathematics (MDPI)
|
Study: White-Box Constructions for ML-KEM (PQ-WB-KEM): a two-construction design space Target: ML-KEM-768 (FIPS 203), 3-share arithmetic masking over Z_q (q = 3329) Author: Uğur Coruh Manifest: experiment_summary.json (regenerated by code/python/run_all.py)
Operand domain. The ML-KEM base multiply operates in the NTT domain, so its secret operand is
shat = NTT(s), which is full-range over Z_q (a CBD-small coefficient-domainsdoes not stay small under the NTT). All measurements below use the real full-rangeshatand two verified lookup-only constructions: A (shared full multiply table) and B (baked per-component tables).
q x q multiply table): 22.16 MB, secret-independent, shared across all components and all k polynomials.q-entry tables.shat (5000 trials, all 128 components, k=3).whitebox_encoded.json): measured protected core A 25.57 MB / B 11.08 MB. A hardened deployment adds estimated shuffling/support tables. The measured base tables are about 11,600× (A) and 33,400× (B) smaller than the 256 GB HFE white-box.Unmasked correct-key correlation ρ ≈ 0.85 (attack succeeds). Three-share masked ρ_max ≈ 0.011 at 10,000 traces — near the 1/√N ≈ 0.01 sampling-noise floor and below the max wrong-key correlation (0.031), so the attack fails (Hamming-weight leakage model, not real hardware traces). Share statistics: chi-square uniformity p ≈ 0.50; share/secret and masked-output correlations (0.037 / 0.023) sit at the noise floor.
On the full-range shat, the plain linear-solve recovers the secret from unencoded intermediates (128/128); against the encoding-unaware (mixing-aware) attacker it fails (0/128) at a measured 3.41 MB encoding cost; the encoding-aware sanity attacker recovers (128/128). This is an empirical negative result for the tested attacker, not a proof — a BGE-style structural recovery of the encodings is open.
Native protected/direct base-multiply ratio (construction B, table + 3-share masking): x86-64 ≈ 4.30×, aarch64/QEMU ≈ 2.85× (functional emulation; ratio is the comparable metric). Embedded deployment is DRAM-bound (7.67 MB > cache; working-set sweep DRAM/L1 ≈ 71× per random access), giving a projected ~47× (B) decapsulation overhead (native 4.30× × ~11× embedded cache; the ~17× figure — ~1.55× RNG-excluded masking floor × ~11× — is the explicitly-labeled lower bound; Construction A is not projected); against the 0.213 ms x86-64 host baseline this is a projected ~10 ms (B) protected decapsulation. Only decapsulation is protected.
| Channel | Estimated security | Notes |
|---|---|---|
| DCA resistance | ~2^12 | 10,000-trace ρ = 0.011 near the noise floor; 1/ρ^2 estimate, not a proof |
| Table footprint | A ~2^24.6 / B ~2^23.4 | Byte-reads to COPY the core: a size, not an extraction cost and not a bound. Table size is NOT a security parameter here – the measured extraction budget is 1,536 B (B) and zero (A), independent of the footprint (entry_determinacy.json). |
| Timing | structural constant-time | no secret-dependent branches; cache-line constant-time not claimed |
Regenerate all results with python code/python/run_all.py (plus the two C benchmarks).