Cloud-Aware Adaptive Defense Framework for MLaaS Against Single and Hybrid Attacks

Abstract

Scalable deployment of machine learning models on a cloud environment on demand, with standardized interfaces has led Machine Learning as a Service (MLaaS) to be a dominant paradigm that has been used for the deployment of machine learning models. But APIs making models accessible represent a major threat to security and MLaaS systems are susceptible to an array of attacks. This paper presents a detailed investigation of single and hybrid attacks against MLaaS systems and introduces an adaptive defense framework based on cloud-awareness to overcome the attacks. The proposed system has features of real-time monitoring, attack detection, and context-aware adaptive defense selection. The dynamic decision making is guided by a mathematical formulation introduced by means of optimization and reinforcement learning. The proposed framework is tested in competitive attack scenarios on CIFAR-10 dataset, with high accuracy rates and an effective reduction of success rates in attacks. The results highlight the importance of adaptable and localized security approaches to protect cloud machine learning systems.