Ensuring security in rapidly developing Optical Networks against sophisticated cyber attacks is paramount. Deep learning (DL)-based detectors analyzing communication logs are widely used for identifying attack-induced faults. However, these initial defense mechanisms are susceptible to adversarial attacks, where malicious actors carefully alter inputs to deceive the detector. This paper proposes a novel, robust attack detector specifically designed to withstand such adversarial manipulations. Our approach is rooted in wavelet insight: adversarial perturbations often cause the detector to incorrectly focus on high-frequency components of the input. By employing wavelet decomposition, our model can analyze the communication logs across multiple frequency scales. This allows us to process or filter specific frequency bands affected by adversarial noise, enabling the detector to focus on robust features that are invariant to such perturbations. We demonstrate through extensive experiments that our wavelet-enhanced detector significantly improves robustness against various adversarial attacks while maintaining high detection accuracy.