Final year Major Project:
In the fields of artificial intelligence (AI) and machine learning, deep learning models have emerged as a new way of learning. The growing interest in Deep Neural Networks is due to their wide applicability in solving day-to-day problems, but there are significant concerns regarding their robustness. This thesis investigates well-known attacks on these models and explores their underlying principles. It also looks into potential defense strategies to counteract these attacks.
A key aspect of this research was to develop a defense method that integrates seamlessly with existing convolutional networks. To this end, the DeepLDA model was implemented, which substitutes the standard categorical cross-entropy layer in the model with a Linear Discriminant Analysis objective function. This study compares the performance of the DeepLDA model against traditional models using categorical cross entropy, particularly in terms of accuracy against adversarial examples.
This research was conducted as part of the Computer Science and Engineering program at the Motilal Nehru National Institute of Technology, Allahabad.