Cooperative Co-evolutionary Genetic Algorithm for Generating Adversarial Examples

Deep neural networks have rapidly advanced, with many models demonstrating outstanding performance. However, robustness is crucial for ensuring the safety and stability of artificial intelligence, especially in real-world applications where models are frequently exposed to interference. Research on adversarial example generation is significant for improving model robustness and preventing sensitive information from being interpreted. This paper proposes a Cooperative Co-evolutionary Genetic Algorithm (CCGA) for generating adversarial examples with small perturbations. The algorithm adopts a collaborative optimization strategy, demonstrating high efficiency when dealing with large-scale problems. Furthermore, it operates as a black-box attack method, requiring only access to the model’s inputs and outputs without any knowledge of its structure or parameters. Experimental results on the MNIST and CIFAR-10 datasets show that CCGA achieves adversarial success rates of 97% and 100%, respectively, with mean squared errors as low as 7.41 and 12.00 between the adversarial and original samples, the differences are nearly imperceptible to human eyes.