Abstract: Evaluating the performance of convolutional neural network models is crucial,and model evaluation serves as a key component in the process,which is widely used in model design,comparison,and application.However,most existing model eva-luation methods rely on running models on test data to obtain evaluation indexes,so these methods are unable to deal with the situation where testing data is difficult to obtain due to privacy,copyright,confidentiality,and other reasons.To address the problem,this paper proposes a novel method to model evaluation that does not require testing data,instead,it is based on feature chirality.The model evaluation obtains the evaluation indexes of the models by calculating the kernel distance of different models.The negative correlation between model performance and kernel distance is then used to analyze model parameters and obtain the relative performance ranking of different models without accessing any testing data.Experimental results show that when using Euclidean distance,the proposed blind evaluation method achieves the highest accuracy across seventeen classic CNNs,including AlexNet,VGGNets,ResNets and EfficientNets.Thus,this method is an effective and viable approach for model evaluation.

Key words: Data-free, Model evaluation, Feature chirality, Distance measure, Convolution kernel distance