Towards Securing Machine Learning Models Against Membership Inference Attacks

From fraud detection to speech recognition, including price prediction, Machine Learning (ML) applications are manifold and can significantly improve different areas. Nevertheless, machine learning models vulnerable exposed security privacy attacks. Hence, these issues should be addressed while using ML preserve the of data used. There is a need secure models, especially in training phase datasets minimise information leakage. In this paper, we present an overview threats vulnerabilities, highlight current progress research works proposing defence techniques against The relevant background for attacks occurring both testing/inferring phases introduced before presenting detailed Membership Inference Attacks (MIA) related countermeasures. introduce countermeasure membership inference on Conventional Neural Networks (CNN) based dropout L2 regularization. Through experimental analysis, demonstrate that technique mitigate risks MIA ensuring acceptable accuracy model. Indeed, CNN model two CIFAR-10 CIFAR-100, empirically verify ability our strategy decrease impact compare results five classifiers. Moreover, solution achieve trade-off between performance mitigation attack.