Provable Privacy Attacks on Trained Shallow Neural Networks