Benford's law describes a regularity in many naturally-occurring numerical datasets: the leading digit does not appear with equal frequency. Election forensics borrowed the test on the hypothesis that fabricated results would struggle to follow the same curve.
This page shows the first- and second-digit distribution of per-section vote counts for each party against the Benford-expected curve. The literature (Mebane) recommends the second-digit test (2BL) over the first-digit test for election data, because per-section vote counts are range-bounded.
This is not evidence of fraud. Plenty of clean electoral data fails the first-digit test. See the full methodology for the nuances.