This thesis investigates the generalizability and decision-making process of a model designed to perturb a modulation classifier under fading channel conditions, with a decision space that includes complex, higher-order modulation schemes. The research aims to evaluate the effectiveness of the perturbation model, its ability to generalize to complex modulations, its impact on communication quality, and the detectability of the perturbations. A perturbation model was tested under fading conditions and across various modulation types. The results demonstrate that the model successfully generates perturbations capable of misleading the eavesdropper, even in the presence of signal impairments due to synchronization issues. However, certain modulation types, specifically 64APSK, proved more challenging to deceive. While communication performance was largely preserved, some perturbations were detectable in the spectral domain for signals with symmetric spectra.The findings indicate that the model can effectively perturb the classifier without significant performance degradation. However, there are opportunities to enhance the stealth of perturbations, particularly for certain modulation schemes. Future work could focus on reducing the detectability of perturbations, exploring the impact of a larger decision space on perturbation power requirements, and incorporating modulation type information more explicitly.