<p>Cancelable biometrics have become a prospective solution in order to deal with privacy and security issues surrounding conventional biometric authentication. Nonetheless, a compromise between good template protection and high recognition performance has not been completely attained. This paper presents an efficient and secure cancelable iris recognition system, which combines hybrid deep feature extraction with Gaussian Random Projection (GRP)-based template transformation. The suggested approach will be based on a combination of MobileNetV3Small with InceptionV3 to obtain both local and global iris features with complementary features, then, feature fusion, feature normalization, and dimensionality reduction will be performed to optimize the discriminative power. GRP is used to convert the fused features into cancelable templates, which are non-invertible, revocable and unlinkable. The suggested framework is tested on several benchmark datasets of irises, such as CASIA-Interval, CASIA-Syn, IIT Delhi, MMU Iris, and a test dataset, which shows high generalization in a wide range of acquisition conditions. Empirical findings demonstrate that the suggested technique results in an accuracy of up to 100 percent with an Equal Error Rate (EER) of 0.0001, which is much higher than the performance of single deep models and hybridized methods in the absence of template protection. Moreover, the framework has minimal computational complexity, thus can be deployed in real-time. Thorough analysis, such as ablation study, security analysis and comparison with state-of-the-art methods verify that the suggested methodology offers effective trade-off of recognition accuracy, computational efficiency, and template security. These findings indicate the feasibility of the given framework in terms of feasible and safe biometric authentication systems.</p>