A compositional framework for automated discovery of adversarial attacks

The vulnerability of deep neural networks to adversarial attacks is a critical barrier to their safe deployment. However, the evaluation of model robustness still relies on a limited set of manually-designed, monolithic attacks, potentially leaving vast regions of the algorithmic search space unexplored. To address this limitation, we introduce a compositional framework that treats adversarial attacks as modular systems synthesized from a grammar of interchangeable functional blocks, including initializers, loss functions, schedulers, and a novel composite gradient computer. This work formalizes the design of new attacks as a black-box optimization problem within this complex, hierarchical space. Metaheuristic strategies, particularly Bayesian Optimization and a custom-designed Evolutionary Algorithm, are employed to automatically discover novel and effective attack configurations. Experiments, conducted on an ensemble of robustly trained ResNet-18 models, demonstrate that attacks discovered through this automated synthesis process provide promising evidence of outperforming established, manually-crafted benchmarks and state-of-the-art transfer baselines (including PGD, MIM, DIM, and the AutoAttack suite) within the evaluated ResNet-18/ImageNette setting, achieving a higher Attack Success Rate under identical perturbation constraints. Our findings suggest that automated compositional synthesis as a promising approach for discovering transferable adversarial vulnerabilities, demonstrating that query-assisted gray-box transfer auditing provides a more robust security evaluation than relying solely on fixed, monolithic algorithms.