Defending Data Poisoning Attack Through Watermarked Friendly Noise Luminosity Activated Dense Layered UNet for Classification of Lung Disease

An accurate and prompt diagnosis is essential since lung illnesses are a major global cause of chronic illness and mortality. Conventional techniques for lung disease diagnosis frequently depend on radiologists manually interpreting medical images, which can be laborious and prone to variability in interpretation. Deep learning methods have been demonstrated recently in automating the diagnosis of lung disorders from medical images, especially CT and chest X-rays. This research proposes Luminosity Activated Dense U-Net (LDU-Net) model inspired by U-Net that classifies the lung disease with high accuracy. This research utilizes the 20,000 chest X-Ray images from COVID-19 Radiography Database for lung disease classification. The proposed LDU-Net is built on Deep Neural Network (DNN), in which Three-Layered Dense Block (3-LDB) serves in each layer of the contrasting and expansion block. The LDU-Net model is initiated by masking the original chest X-Ray image with its respective lung segments to form the masked chest X-Ray image. To enhance the lung segmentation, the essential features from masked chest X-Ray images are validated with the luminous intensity through magnitude histogram mapping of image pixels. The masked chest X-Ray images are subjected to form exponential transformed, absolute differenced, contrast and luminous image. The image intensity ratio observed to be high with Luminous Masked Chest X-Ray Vector (LMC) image. To address the data poisoning attack towards security countermeasure, the LMC images are processed to form Watermarked Friendly Poison LMC (WFLMC) image by the proposed Watermarked Friendly Noise Luminosity Activated Dense U-Net (WFLDU-Net) model. The WFLMC images are fitted with 3-LDB in each of three contrasting blocks of the LDU-Net that down samples the feature map by factor 2 and gradually preserves the feature map latent representation by learning the useful features. Each contrasting block is associated with the dropout that alleviates in learning the most essential features from WFLMC images. Then the WFLMC feature maps from the contrasting block are fitted with three expansion blocks that up sample the feature map by factor 2. With the presence of skip connections copy and copy, the decoder block learns the spatial information from the up sampled WFLMC image by replicating the segmentation mask. The novelty of the proposed LDU-Net is attained by integrating the 3-LDB with the proposed U-Net which is again integrated with security countermeasures. The adversarial attempt in the form of data poisoning attack is defended by the proposed WFLDU-Net model with the creation of WFLMC images. The implementation results reveal the proposed LDU-Net model has been found to exhibit high accuracy of 99.36% towards lung disease classification.