ATTACKING AND PROTECTING DATA PRIVACY IN EDGE-CLOUD COLLABORATIVE INFERENCE SYSTEMS

Abstract

Intelligent applications are being developed almost at geometric speed. It is such that the fast-growing edge-cloud collaborative paradigm allows harnessing cloud computational power while reducing latency through the edge devices. However, this paradigm introduces great privacy challenges in model partitioning, transmission, and joint inference. The sensitive user data processed at the edge and shared with the cloud can be vulnerable to poisoning and attacks like model inversion, adversarial perturbations, and membership inference, which can cause glaring privacy issues. In combating these attacks and threats, it is pertinent to steer the strategy towards a balance between security, efficiency, and inference accuracy. The paper will explore data privacy-related attacks and defenses in the collaborative edge-cloud inference systems. It looks into possible attack scenarios that exploit loopholes existing at both the edge and at the cloud, while simultaneously enlightening scenarios where an adversary could extract or manipulate sensitive information. Furthermore, it presents a discussion on cutting-edge protections, including differential privacy, homomorphic encryption, secure multi-party computation, and adversarial defenses.