Título: A Dataless Approach to Latent Evolutionary Images
Autores: Caio Santana Trigueiro, Arthur Buzelin, Pedro Bento, Yan Aquino, Victoria Estanislau, Samira Malaquias, Wagner Meira Junior & Gisele Lobo Pappa
Resumo: Building on our previous Evolutionary Bias Identification via Embeddings (EBIE) framework, we present Latent Evolutionary Images (LEI), a dataless framework for auditing black-box image classifiers by evolving interpretable synthetic images in the latent space of a generative model. Unlike traditional adversarial or interpretability approaches, LEI operates entirely without labeled training data or gradient access, relying instead on a genetic algorithm that explores the latent space of a pre-trained Variational Autoencoder (V AE). A frozen classifier evaluates the fitness of each generated image, steering evolution toward samples that maximize confidence in a chosen target class. We demonstrate that this process reliably uncovers latent directions associated with class semantics, even in the absence of explicit supervision. Experimental results show that LEI not only produces visually coherent representations aligned with the classifier’s internal logic, but also reveals class-specific biases and facilitates the generation of effective latent adversarial examples1.
Palavras-chave: evolutionary algorithms; latent space optimization; variational autoencoder; black-box model auditing; image classification; image generation.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1191894
Artigo em PDF: CBIC_2025_paper1191894.pdf
Arquivo BibTeX:
CBIC_2025_1191894.bib