On Manipulability of Algorithms

In adversarial situations, the input data to an algorithm could be manipulated to make the algorithm produce erroneous output or to make a wrong decision. The paper presents a formal definition and a model of algorithm manipulation from a game-theoretic point of view. Algorithm manipulation is viewed as a game between a decision maker and an adversary. The decision maker runs an algorithm to make a decision, whereas the adversary manipulates the input data to his own advantage and to the disadvantage of the decision maker. The paper also proposes a method for decision making based on manipulated input. According to the method, the decision strategy and the manipulation must be in Nash equilibrium. In other words, the decision strategy is the best response to the manipulation and vice versa, the manipulation is the best response to the decision strategy.