To defend against denial-of-service (DoS) attacks, we employ a technique called resource burning (RB). RB is the verifiable expenditure of a resource, such as computational power, required from clients before receiving service from the server. To the best of our knowledge, we present the first DoS defense algorithms where the algorithmic cost -- the cost to both the server and the honest clients -- is bounded as a function of the attacker's cost. We model an omniscient, Byzantine attacker, and a server with access to an estimator that estimates the number of jobs from honest clients in any time interval. We examine two communication models: an idealized zero-latency model and a partially synchronous model. Notably, our algorithms for both models have asymptotically lower costs than the attacker's, as the attacker's costs grow large. Both algorithms use a simple rule to set required RB fees per job. We assume no prior knowledge of the number of jobs, the adversary's costs, or even the estimator's accuracy. However, these quantities parameterize the algorithms' costs. We also prove a lower bound on the cost of any randomized algorithm. This lower bound shows that our algorithms achieve asymptotically tight costs as the number of jobs grows unbounded, whenever the estimator output is accurate to within a constant factor.
翻译:暂无翻译