Expanding vzn's comment into an answer: The standard reduction from CNF-SAT to vertex cover is pretty easy: make a vertex for each term (variable or its negation), connect each variable to its negation by an edge, make a clique for each clause, and connect each vertex in the clique to the vertex for one of the terms in the clause. If you start with a satisfiability problem with a known satisfying assignment, this will give you a vertex cover problem with a known optimal solution (choose the term vertices given by the assignment, and in each clause clique choose all but one vertex, so that the clause vertex that is not chosen is adjacent to a term vertex that is chosen).
So now you need to find satisfiability problems that have a known satisfying assignment but where the solution is hard to find. There are many known ways of generating hard satisfiability problems (e.g. generate random k-SAT instances close to the satisfiability threshold) but the extra requirement that you know the satisfying assignment restricts the possibilities. One thing you can do here is go through another level of reduction, from a cryptographically hard problem such as factorization. I.e. choose two large primes p and q, set up a Boolean circuit for multiplying p and q as binary numbers, and translate it into a CNF formula in which there is a variable for each input (p and q) and for each intermediate value on a wire in the circuit, a clause for each output forcing it to have the right value, and a clause for each gate forcing the inputs and outputs of the gate to be consistent with each other. Then translate this CNF formula into vertex cover.
For a simpler strategy, choose the satisfying assignment to a 3CNF formula first, and then generate clauses at random, keeping only the clauses that are consistent with the assignment, and then convert to vertex cover. If the clauses have uniform probability this will be vulnerable to a degree-based heuristic (the term vertices that match the chosen assignment will have lower degree than the term vertices that do not) but this shortcoming can be avoided by adjusting the probabilities of the clauses according to how many of the clause's terms agree with the chosen assignment. Probably this is vulnerable to some kind of polynomial time attack but it might not be one that's natural for vertex cover, so it might make a good set of test instances despite not having much guarantee of hardness.