Enhanced Decision Accuracy of Artificial Intelligence Systems using Protection against Cropping Loss Attack by Encryption and Median Filtering

Abstract

Concerns regarding artificial intelligence systems' susceptibility to hostile assaults have grown in recent years due to their growing integration in a variety of applications. Cropping loss attacks are one particularly dangerous form of cyber-attack, in which attackers alter precise areas of the input image to trick AI algorithms into drawing false conclusions. In order to protect artificial intelligence models from cropping loss attacks, this research proposes a novel method that combines encryption with a noisy pixel centered median filter. To assess the effectiveness of our suggested strategy, in-depth tests carried out utilizing the kaggle’s cats and dogs image set and convolutional neural network VGG19.  The results showcase a significant improvement in the resilience of AI systems against cropping loss attacks. Mean square error improvement factor greater than 117.8219 for large cropping up to 52.94% region of an image without median filtering and with median filtering is 500.2004. The median filter and encryption work together to improve artificial intelligence models' overall decision-making accuracy while also thwarting adversary manipulations. These results are applicable to a variety of domains, including as face recognition, object detection, image categorization, and medical diagnostics. The suggested approach offers a strong resistance against cropping loss assaults, advancing the creation of artificial intelligence systems that are more reliable.