Abstract
Malaria remains a serious global challenge that affects millions of people. A bite from female Anopheles mosquitos transfers the disease to human hosts. Early diagnosis and treatment are essential for preventing severe illness and death. The current diagnostic techniques the World Health Organisation (WHO) recommends are microscopy and Rapid Diagnostic Tests (RDTs). These require trained experts and are time-consuming and prone to human error. Deep learning models have shown great promise for automating malaria diagnosis from blood smear images. However, these models are vulnerable to adversarial attacks and, therefore, not robust enough to be deployed commercially. This dissertation addresses this pressing issue by enhancing the robustness of malaria medical image classification using a combination of data augmentation and adversarial training techniques. Three models are compared by accuracy metric and efficiency: baseline model, model trained with augmented data and adversarially trained model. Our findings suggest that both techniques can effectively enhance the robustness of malaria classification models. Adversarial training is the more efficient technique, with an accuracy improvement of 38% from the baseline model. The dissertation contributes to the medical image classification field and paves the way for deploying malaria classification models.