We analyze uncertainty relations on finite dimensional Hilbert spaces expressed in terms of classical fidelity, which are stronger then metric uncertainty relations introduced by Fawzi, Hayden and Sen. We establish validity of fidelity uncertainty relations for random unitary matrices with optimal parameters (up to universal constants) which improves upon known results for the weaker notion of metric uncertainty. This result is then applied to locking classical information in quantum states and allows to obtain optimal locking in Hellinger distance, improving upon previous results on locking in the total variation distance, both by strengthening the metric used and by improving the dependence on parameters. We also show that general probabilistic estimates behind the main theorem can be used to prove existence of data hiding schemes with Bayesian type guarantees. As a byproduct of our approach we obtain existence of almost Euclidean subspaces of the matrix spaces $\ell_1^n(\ell_2^m)$ with a better dimension/distortion dependence than allowed in previously known constructions.