부산대학교 도서관

The performance of distributed streaming systems relies on how even the workload is distributed among their machines. However, data and query workloads are skewed and change rapidly. Therefore, multiple adaptive load-balancing mechanisms have been proposed in the literature to rebalance distributed streaming systems according to the changes in their workloads. This paper introduces a novel attack model that targets adaptive load-balancing mechanisms of distributed streaming systems. The attack reduces the throughput and the availability of the system by making it stay in a continuous state of rebalancing. This paper proposes Guard , a component that detects and blocks attacks that target the adaptive load balancing of distributed streaming systems. Guard uses an unsupervised machine-learning technique to detect malicious users that are involved in the attack. Guard does not block any user unless it detects that the user is malicious. Guard does not depend on a specific application. Experimental evaluation for a high-intensity attack illustrates that Guard improves the throughput and the availability of the system by 85% and 86%, respectively. Moreover, Guard improves the minimum availability that the attacker achieves by 325%.