Adversarial-Aware Population Search for Cybersecurity Decision Tuning with a Finite Attack Surrogate

Abstract

Cybersecurity decision thresholds and ensemble weights can be tuned under an adversarial loss, but doing so requires an explicit attacker and evaluation protocol. This manuscript describes Adversarial-Aware Population Optimization (AAPO), an outer population search using differential-evolution, particle-swarm and Lévy-style updates. Its supplied pseudocode evaluates each candidate through an average of projected-gradient attack losses across batches and restarts. That finite surrogate is below the exact constrained worst-case loss on the same inputs; it is not a certificate for the stated min–max objective or a demonstrated query-limited grey-box adversary. In the supplied MNIST summaries at an L∞ budget of 0.1, robust accuracy is 80.1% for AAPO and 74.2% for the Madry-PGD-trained comparator, a difference of 5.9 percentage points. Reported pairwise statistics and sensitivity values are retained as source observations, but their test units, raw ranks and execution inputs require reconciliation. In particular, the seven published mean ranks total 29.90 rather than 28, and the stated Nemenyi critical difference of 1.94 does not follow from seven methods and eight independent blocks. Source figures and tables also disagree on several comparator accuracies. The method is consequently presented as a proposed decision-layer tuning procedure with preliminary, unverified benchmark summaries; publication-level superiority, certified L∞ robustness and optimizer guarantees await code, data and corrected statistical outputs.