Using Convolutional Network Visualisation to Determine the Most Significant Pixels

TL;DR: This paper shows how to combine Class Activation Map with feature map to determine a few of the most contributing pixels for given input and modify them to perform an adversarial attack.

Abstract: Over the last years, Deep Neural Network models have been recognized as successful in solving many complex problems. However, these methods are mostly focused on the efficiency of final results and rarely provide sufficient evidence and details on factors that contribute to their outcomes This is why a growing demand for analysis techniques appeared. Thanks to visualisation techniques we can if network works as expected or even improve output of given model if possible. Moreover we can use these methods as optimization technique to boost network’s performance but pruning less important neurons. Finally, if we know how a given model works we can prepare a disruption to its work process. This paper shows how we can combine Class Activation Map with feature map to determine a few of the most contributing pixels for given input and modify them to perform an adversarial attack.