An exhaustive analysis of Test Vector Leakage Assessment (TVLA) methodologies to detect side-channel information leakage in constant-time ML-KEM hardware and software engines.
Non-specific Test Vector Leakage Assessment utilizes Welch's t-test to evaluate whether two sets of power consumption or electromagnetic emission traces—one generated with fixed inputs and one with random inputs—exhibit statistically significant differences. A t-statistic exceeding |t| = 4.5 indicates information leakage with 99.999% confidence.
In ML-KEM, the Number Theoretic Transform (NTT) and Montgomery modular reduction are prime targets for side-channel scrutiny. Memory lookups indexed by secret coefficient data or conditional subtraction instructions in reduction routines generate distinct power signatures. TVLA pinpoints the exact clock cycle of vulnerability.
To achieve compliance below the |t| < 4.5 threshold across millions of traces, hardware engines employ higher-order arithmetic masking, randomized polynomial coefficient processing order, and dummy cycle insertion, neutralising power and EM side-channel vectors.