Photographs capture our feelings, our moments, and as human beings, we love to remember and share moments. With the rapid growth of social media websites and apps like Facebook, Whatsapp, Instagram and Snapchat, people are sharing their photographs and images like never before. Also, the technological advancements have made devices to capture photographs accessible to everyone. This has created a vast amount of digital photographs, leading to the requirement for new technologies to be developed to efficiently archive and access this vast collection of photographs.Text based search and retrieval has become very efficient with modern search engines like Google and Bing. Naturally, similar technologies can be applied to image based tasks if an image can be explained using text by either assigning a few tags, a short caption, or a full textual explanation. Assigning relevant tags to an image is called image annotation. Image annotation is a promising technique for indexing and searching large collections of images. In this context, diverse image annotation refers to predicting tags which are representative and diverse. Since it is not feasible to do manual annotation, automated computational mechanisms have become necessary to perform such tasks.Several attempts have been made to address the problem of automatic image annotation. Initially people targeted it by translating an image into a few keywords and by mapping a relevance between features of an image and tags. Lately,advancements in deep learning has helped in significantly improving the performance of image annotation tasks. In parallel, recent research has demonstrated the impact of adversarial examples, which act as optical illusions for machine learning algorithms.These examples are slight modifications to the original image and very hard to distinguish by human eye, and are produced intentionally to fool machine learning models.In this thesis, we first propose a data-independent noise computing mechanism.It computes the worst case perturbations and adds it in test features to generate adversarial samples, assuming that the training data is not available to the attacker.In this set-up, the performance of eight state-of-the-art multi-label prediction algorithms against worst-case perturbations has been studied on two benchmark datasets. After this, we analyze the effect of feature normalization on robustness of these algorithms. Extensive experiments show that feature normalization does help to improve the performance of some of the algorithms, thus opening-up a new direction for further analyzing the impact of this simple technique in building robust machine learning models.