Enhancing Data Privacy and Governance in Cloud-Deployed Large Language Models Using Deep Learning-Based Risk Mitigation

Abstract

The growing number of problems related to the privacy and security of data in cloud computing requires the need for efficient protection schemes. In this paper, we present an innovative solution for ensuring the privacy and management of data in large language models deployed in clouds by applying risk mitigation techniques based on deep learning. The technique involves risk mitigation, secure data management, and model performance enhancement. First, data input is acquired from a synthesized cybersecurity logs dataset and filtered using data preprocessing to eliminate any form of inconsistencies. Next, data preprocessing enables the input to be directed to the feature extraction phase utilizing the Dense Channel Spatial Semantic Guidance Attention UNet (DCSGA-UNet), which allows for the extraction of relevant features, such as Log_ID, IP_Address, Request_Type, and Response_Time_ms. After that, these features undergo processing through a Bayesian Constitutive Artificial Neural Network (BCANN) classifier that helps predict the probability of a threat or non-threat occurrence. The technique is designed using Python and evaluated against current methods. The experiments show that our innovative solution exhibits a higher accuracy of 98% in a lesser amount of time of 1.07 seconds.