Evolving binary masks to explain image classification

The high dimensionality of image tensors and the complex architecture of Deep Convolutional Neural Networks render the task of image classification very non-intuitive, prompting the need for tools like eXplainable Artificial Intelligence(XAI) to comprehend the classification strategy. While notable techniques such as Gradient Weighted Class Activation Mapping, SHapely Additive exPlanations, and Local Interpretable Model-Agnostic Explanations have made significant strides in this domain, there remain some challenges concerning their applicability in scenarios where the model is entirely abstracted from the user and leveraging those explanations for understanding the classification strategy of the networks as inconsistencies prevail between the predictions for the original image and its corresponding explanations. A novel model-agnostic XAI algorithm with an evolutionary approach is introduced in this paper to provide local explanations for image classification tasks that address the drawbacks encountered by the mentioned algorithms. Central to this approach is the evolution of Binary Masks, which is then used to explain the classification strategy of the network. Sample results with benchmark datasets like CIFAR 10 and Imagenette have been documented along with a comparative study for the explanations obtained from the state-of-the-art techniques against the proposed algorithm. The results reveal the efficiency of the proposed algorithm against the existing algorithms in terms of the confidence and localization capabilities for the corresponding explanations.