Interpretabilidade por otimização: uma análise da geração de explicações black-box com algoritmos genéticos

Abstract

With the advancement of artificial intelligence and the use of deep learning models, it has become essential to understand why and how these models arrive at their results, ensuring their reliability in critical applications. Model interpretability refers to the ability to explain the internal behavior and decisions of these systems, making them transparent and auditable. However, many existing explainability methods, especially those based on gradients, often produce noisy, fragmented, and inconsistent visualizations, limiting their practical usefulness and user confidence. Given this gap, this work proposes an optimization-based approach to improve the interpretability of computer vision models, aiming to generate more intuitive and reliable explanations for humans. Specifically, we propose and validate the IPHA-GA method, which uses genetic algorithms to treat interpretability as a black-box optimization problem, seeking to generate consistent and high-fidelity explanations. The method was systematically evaluated through six main experiments. First, IPHA-GA was validated on image classification tasks (ResNet-18 on ImageNet), where it outperformed traditional gradient-based methods (Grad-CAM, Integrated Gradients, Saliency Maps, Guided Backpropagation) in identifying both relevant pixels (Feature Impact Index ≤ 8%) and irrelevant pixels (FII ≥ 93%). Second, an ensemble approach was developed that uses traditional methods as the initial population for IPHA-GA, achieving up to 22% gains in FII. Third, the method was extended to vision-language models (CLIP), exploring three mask granularities (pixel, patch, region of interest), where the ROI approach balanced quantitative accuracy with human interpretability. Fourth, a stability analysis with 30 different seeds demonstrated high reliability of the explanations (mean variance 2 = 0.032). Fifth, contextual bias was investigated across multiple architectures (YOLO11, CLIP), revealing that model complexity does not correlate with background dependence (∼42% IoU for YOLO, ∼32% for CLIP), but rather with characteristics of the training data. Sixth, the GA-Attacker framework was developed to assess adversarial fragility, demonstrating that optimization-based methods, as well as gradient-based ones, are vulnerable to imperceptible perturbations (IoU dissimilarity = 44.6%). We conclude that the formulation via genetic optimization is promising for interpretability, offering high fidelity at the cost of increased computational demand, and being particularly valuable for critical applications such as medical diagnosis and autonomous systems.