Feature selection is a well-studied problem, yet established criteria—variance (PCA), statistical dependency (mutual information), neighbour distance (ReliefF), and recursive elimination (RFE)—are all blind to the adversarial threat model: they optimise for representational fidelity, not for robustness against malicious perturbation. This paper introduces QEFFE (Quantum Entropy-based Feature Forward Elimination), a feature-selection method whose selection criterion is the symmetric Kullback–Leibler divergence between a feature\'s clean-input and adversarial-input activation distributions, combined with a redundancy penalty inside a greedy forward-selection loop. The “quantum” label denotes an entropy-divergence design metaphor and not quantum hardware. Evaluated on a 512-dimensional deep-residual feature space derived from MNIST, QEFFE attains the highest dimensionality reduction among five compared methods (81.25%, 512?96) and, as an isolated component, raises projected-gradient-descent (PGD, ?=0.20) accuracy by 33.0 percentage points—the single largest contributor in a full-pipeline ablation. The method adds zero inference-time parameters, operating as an index lookup. Results on EMNIST Balanced (47 classes) confirm that the criterion transfers beyond the binary-scale digit task.