ABSTRACT
This paper introduces Adaptive NERO-Net, an extension of the neuroevolutionary framework, to improve search efficiency and robustness in CNNs. It includes an adaptive mutation strategy, a multi-objective fitness function, and comprehensive evaluation across datasets. The approach achieves superior robust accuracy and reduces search generations.
PAPER · PDF
Loading PDF...
Key findings
Adaptive NERO-Net improves search efficiency and robustness in CNNs.
Achieves 52.8% robust accuracy under PGD-20 attacks on CIFAR-10 with only 2.8M parameters.
Reduces search generations by 30% compared to fixed hyperparameters.
Limitations & open questions
Further research needed to generalize findings to other datasets and attack types.
Computational cost of fitness evaluation under adversarial attacks remains high.