Enhanced IRIS Authentication Model - A Gradual Feature Selection Reduction Approach Augmented with Soft Vote Ensemble

The issues identified in primary studies showed the need for a reliable user authentication system and thereby constitute low performance metrics. This include the restriction of feature extraction to single image filter thereby returning relative feature vector size of biological traits returns sub-optimal models for a data mining-based authentication system. The aim of this study is to enhance the performance of existing models deploy SqueezeNet embedding for image feature extraction as well as the implementation of synthetic oversampling and employment of vote ensemble model to address the widespread problem of under-fitting and over-fitting observed in literatures. The study further implemented feature engineering technique of feature selection (through information gain and Relief-F) to enhance model performance by eliminating feature redundancy. The user authentication was actualized through Vote ensemble modelling of SMO, MLP, and the DT base learners. Results from the study showed a precision of 75.37%, an accuracy of 83.7%, F-Measure value of 0.781917211 and ROC AUC value of 0.833333333 indicating that the model has a good capability to differentiate between the positive and negative classes.