Implementation of impersonation attacks against machine learning-based biometric authentication systems

The research delves into the evolving field of iris recognition in biometric systems, emphasising the dangers of reconstruction attacks. It addresses the challenge of maintaining these systems’ security and reliability in the face of sophisticated attack strategies. The CASIA V1.0 dataset with 756 iris images was used in the study to build and test iris recognition models. Two primary models were created, one inspired by literature. Convolutional neural networks were used in these models, which were rigorously trained and validated. The project also investigated various reconstruction attack strategies, with a particular focus on reconstructing training data from released machine learning models using a reconstructor network. The recognition models achieved high iris identification accuracy, with the first model achieving 96.43% and the second 92.86%. The reconstruction attack experiments, on the other hand, revealed significant differences in the biometric similarity of the reconstructed and actual iris images. To assess these differences, descriptive statistics and statistical analyses (including the Shapiro-Wilk test and paired t-tests) were used. The study demonstrates that iris recognition systems can maintain a high level of security and data integrity even when subjected to sophisticated reconstruction attacks. The similarity between the reconstructed and actual iris images suggests that these systems are resistant to model inversion attacks, which increases trust in biometric security systems. The study concludes that modern iris recognition models are not only highly accurate, but also have a high resistance to reconstruction attacks. This highlights the importance of these systems in secure biometric verification and identification processes, providing strong defence against potential security threats