This study investigates how the structural and spectral properties of Boolean functions influence their susceptibility to machine learning–based cryptanalysis. A comprehensive framework is proposed, combining Boolean function generation, truth table representation, Walsh–Hadamard spectral analysis, and machine learning evaluation. Using a dataset of 2000 functions (linear, balanced, bent-like, and random), results show that higher nonlinearity reduces learnability, while linear functions remain predictable. A strong negative correlation (−0.6499) between nonlinearity and learning accuracy is observed. Functions with large Walsh coefficients are more easily approximated. The findings confirm that machine learning exploits inherent structural weaknesses, aiding the design of more secure, learning-resistant cryptographic primitives.