FLDP: A differentiated privacy security approach based on flow model

Sensitive data in standard deep learning training can be subject to assaults. These attacks take advantage of the disclosure of gradients and intermediate representations during the training process, potentially resulting in data breaches and unauthorized access to sensitive information. Concerns about the privacy and security of sensitive data during the training process have grown in importance as deep learning has grown in popularity. This study describes a novel strategy FLDP for improving the privacy and security of deep learning training by combining differential privacy (DP) with stochastic gradient descent (SGD) and flow model. Our strategy improves privacy and security without compromising deep learning model training efficiency or scalability. Individual data points are indistinguishable in our methodology, reducing the danger of privacy breaches.Instead of applying a consistent privacy level to all data points, our technique enables for privacy protection to be differentiated based on data sensitivity. This adaptive privacy approach promotes value while preserving the most sensitive data's privacy.