Generating Adversarial Examples through Latent Space Exploration of Generative Adversarial Networks
L. Clare, João Correia · 2023
Artificial Neural Networks are vulnerable to adversarial examples, malicious inputs that aim to subvert neural networks' outputs. Generative Adversarial Networks (GANs) are generative models capable of generating data that follows the training data distribution. We explore the hypothesis of using the latent space of the trained GAN to find adversarial examples. We test the adversarial examples on external classifiers trained on the same training data. Thus, we propose a framework for Generating adversariaL exAmpleS through latent Space Exploration (GLASSE). A Genetic Algorithm evolves latent vectors as individuals and uses a trained GAN to generate examples to maximise a target activation value of the discriminator network. After the evolutionary process, an external classifier trained on the same dataset evaluates whether it is adversarial. The results indicate that we can optimise the objective and find adversarial examples. We tested the generated examples with models from the adversarial learning literature, showing that 82% on average of the generated examples resulted in successful attacks. We show a t-SNE analysis of the examples, showcasing that generated adversarial examples are blended in the cluster of each belonging class and visually similar to the training dataset examples, showcasing the viability of the proposed approach.