TVLA — the Test Vector Leakage Assessment — is the standard tool for detecting side-channel leakage in hardware. It applies Welch's t-test: compare power traces from two groups (different inputs or different keys) and check whether the means differ. If they do, the device leaks information. If they don't, it passes.
Mikulec, Breier, and Hou show that passing TVLA does not mean passing. The t-test detects only mean-based differences. Leakage that manifests as variance changes, distributional shape shifts, or tail behavior differences — anything that alters the distribution without moving the mean — is invisible to TVLA.
The Anderson-Darling Leakage Assessment (ADLA) tests equality of the full cumulative distribution functions. Instead of asking “do these two groups have the same average power consumption?”, it asks “do these two groups have the same distribution?” The distinction matters when protections like shuffling and random jitter are applied — countermeasures that are specifically designed to obscure mean differences while potentially leaving distributional signatures intact.
On a neural network implementation with shuffling and jitter protections, ADLA detects leakage with fewer traces than TVLA. The protected implementation that passes the standard test fails the distributional test. The leakage was always there. The measurement tool was looking at the wrong statistic.
The structural insight: a security test that examines only one property of the signal (the mean) can be satisfied by countermeasures that suppress only that property. The leakage migrates from the tested statistic to untested ones. The countermeasure does not eliminate information flow — it moves it from where you are looking to where you are not.