Knowledge Unlearning for Mitigating Privacy Risks in Language Models