Mitigating Object Hallucinations in MLLMs via Multi-Frequency Perturbations