Avaliação comparativa de redes neurais: na classificação de imagens reais e sintéticas

Abstract

This work investigates the detection of images generated by artificial intelligence by comparing two modern approaches: a detector based on YOLO and a classifier based on CLIP L-14. The motivation arises from the increase in synthetic content and the need for solutions that reconcile accuracy, sensitivity, and computational costs in practical scenarios. The objective is to empirically compare the performance and limitations of both approaches in the task of classifying whether an image is real or fake, identifying trade-offs between accuracy, recall, F1, and computational cost. YOLO is trained at scale with 100,000 training images and 20,000 validation images, performing three experiments with 10, 20, and 30 epochs. CLIP L-14 is initially evaluated in zero-shot mode and then subjected to fine-tuning with the AdamW optimizer in two experiments of 10 and 20 epochs. YOLO achieved the best overall balance in the 10-epoch experiment, with higher F1 and superior accuracy, while longer experiments increase recall at the cost of a higher false positive rate. In contrast, CLIP in zero-shot mode showed a bias toward classifying images as fake and low accuracy, but fine-tuning with AdamW substantially improves the metrics, achieving competitive results in precision and F1. In the adopted experimental configuration, YOLO demonstrates the most efficient solution in terms of performance and com putational cost, whereas CLIP shows potential after supervised adjustment, conditioned on the vailability of infrastructure for training at a comparable scale.