Mitigating Data Integrity Attacks in Building Automation Systems using Denoising Autoencoders

Abstract

Building automation systems (BAS) are a class of cyber-physical systems that aim to improve the efficiency of buildings through intelligent control. Typically, sensor measurements are used to compute for the optimal control action for equipment that minimizes the energy consumption while maintaining the comfort of occupants. Given the heavy reliance on sensor data, ensuring their validity is an essential concern in BAS. Several work have explored sensor validation, mostly for the purpose of automatic fault detection and diagnostics (AFDD). However, recent studies show BAS vulnerable to adversarial attacks - an area that has lacked consideration in the early design of BAS. In this work, we propose a sensor correction model, based on deep learning framework, for mitigating against data integrity attacks in BAS. We test our approach on real-world data obtained from a retrofitted air conditioning (AC) control testbed, with injected data simulating different attacker types and number of attacked sensors. We show that the model is capable of mitigating attacks by comparing with a baseline where no model is employed.