Iris recognition is one of the automated processes of verifying individuals' identity based on their iris characteristics. Apparently, the random nature of the iris texture, which is unique for each individual, makes it an exclusive trait for biometric recognition even for the case of identical twins' authentication. Recently, the improvement in deep learning and computer vision indicated that the extracted features using convolutional neural networks (CNNs) are suitable to describe the complex image patterns. But, how to protect the biometric data and provide users' privacy is a main concern, nowadays. In this paper, we study the performance of pre-trained CNNs to successfully classify cancelable iris features when taking the feature vector from each fully connected layer. We show that these pre-trained CNNs, while originally learned for classifying generic objects, are also extremely good for representing iris images for recognition. The performance metrics are evaluated on three datasets: CASIA-IrisV3, IITD and Palacky iris databases. The obtained results achieve promising cancelable iris recognition and also ensure the robustness and effectiveness of the proposed approach.