The widespread adoption of machine and deep learning techniques for Android malware detection has driven attackers to develop sophisticated evasion strategies, particularly using Generative Adversarial Networks (GANs). This study evaluates the robustness of popular ML-based classifiers against adversarial Android malware samples generated by a GAN. The experiments are conducted under two configurations—overlapped vs. isolated datasets—to analyze the impact of shared versus separate training data between the classifiers and the GAN. Results indicate that classifiers experience a significant decline in performance when adversarial samples are introduced without explicit exposure during training, with misclassification rates reaching 100% on test sets. However, retraining the classifiers with a mix of adversarial samples and a subset of unmodified malware significantly improves resilience, consistently achieving over 95% accuracy for both original and adversarial malware. Additionally, implementing a multi-class classification approach (benign vs. original vs. adversarial) enables models to effectively detect manipulated samples, resulting in high adversarial sample detection rates.