A multiangle evaluation method for target detection models considering adversarial sample attack factors

Le résumé fourni par la source

When evaluating neural network target detection models based on traditional methods, most of the models are evaluated from a single perspective, such as the recognition accuracy or robustness of the model, and the evaluation conclusions have certain limitations. The article proposes to take the adversarial sample attack in the field of artificial intelligence security into consideration, and proposes to comprehensively evaluate the model from multiple perspectives, such as the recognition accuracy, robustness and recognition reliability of the model, firstly, constructing three types of test samples, such as unfamiliar samples, transformed samples, and adversarial samples, and then, through the statistical comparison and analysis of the prediction results and the labels of the test samples, combined with the indexes of recognition correctness, misdetection rate, and leakage rate, the model can finally be evaluated quantitatively. Finally, the model can be evaluated quantitatively, and give the model evaluation conclusions including four graded grades: fail, pass, good and excellent. This model evaluation method can provide certain reference value for the rational use of the model.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.