Project 4: Security, Privacy and Resource Trade-offs

Description:Given anomaly detection requires contextual information (e.g., smart meter/energy pricing data, or smart vehicle trajectory), how to ensure computations and analytics that preserve end user’s privacy? How to provide security and privacy for resource constrained IoT and edge computing applications supporting CPS [PGN+20, TTY+21]? We have recently adopted fully homomorphic encryption (FHE) coupled with fast anomaly detection algorithm and trustworthy decisions in smart grid [IBY+20]. We have also defined new information-theoretic metrics for trade-off analysis between resource utilization and responsiveness (timeliness) vs. security, privacy, and trustworthiness in CPS decision support systems. Moreover, the huge amount of data analyzed for predictions through adversarial machine learning to aid in the intelligent decision making, is vulnerable to manipulation, leading to wrong inference or learning when smart adversaries inject false data intelligently during the training phase. Although machine learning offers a powerful tool to detect and localize security breaches [DBD20, DBD+, LRS+20], traditional algorithms are often complex and mostly used for offline predictions or non-real time decision making. We are investigating novel machine learning methods that are immune to the training set manipulation. For example, in [BTS+21, RSD20], we proposed a light-weight machine learning technique to secure smart grid against false data injection attacks. Students will extend it to investigate advanced data sanitization and robust statistical techniques to secure CPS under adversarial training manipulation and perform experiments to lean insight about the impact of attacks.

Sample Experiment: Performance evaluation of fully homomorphic encryption-based anomaly detection.

Purpose: To train students how to analyze security-privacy tradeoff under resource constraints.

Method: The students will learn how to implement encryption in anomaly detection technique for smart grid application, run the code in sensor cloud, and analyze energy consumer’s privacy and security trade-off.

Input Parameters: Real traces data files, parameters for models to inject false data.

Output Parameters: Estimation of security and privacy breaches; computation cost of proposed solution.

Project Deliverables: Experimental results; modification to algorithm and implementation; and a research report, possibly leading to a conference publication.

Researcher