Catálogo de publicaciones - libros

Graph games played by two players over finite-state graphs are central in many problems in computer science. In particular, graph games with -regular winning conditions, specified as parity objectives, which can express properties such as safety, liveness, fairness, are the basic framework for verification and synthesis of reactive systems. The decisions for a player at various states of the graph game are represented as strategies. While the algorithmic problem for solving graph games with parity objectives has been widely studied, the most prominent data-structure for strategy representation in graph games has been binary decision diagrams (BDDs). However, due to the bit-level representation, BDDs do not retain the inherent flavor of the decisions of strategies, and are notoriously hard to minimize to obtain succinct representation. In this work we propose decision trees for strategy representation in graph games. Decision trees retain the flavor of decisions of strategies and allow entropy-based minimization to obtain succinct trees. However, decision trees work in settings (e.g., probabilistic models) where errors are allowed, and overfitting of data is typically avoided. In contrast, for strategies in graph games no error is allowed, and the decision tree must represent the entire strategy. We develop new techniques to extend decision trees to overcome the above obstacles, while retaining the entropy-based techniques to obtain succinct trees. We have implemented our techniques to extend the existing decision tree solvers. We present experimental results for problems in reactive synthesis to show that decision trees provide a much more efficient data-structure for strategy representation as compared to BDDs.

Despite the improved accuracy of deep neural networks, the discovery of adversarial examples has raised serious safety concerns. Most existing approaches for crafting adversarial examples necessitate some knowledge (architecture, parameters, etc) of the network at hand. In this paper, we focus on image classifiers and propose a black-box approach to test the safety of deep neural networks that requires no such knowledge. Our algorithm employs object detection techniques such as SIFT (Scale Invariant Feature Transform) to extract features from an image. These features are converted into a mutable saliency distribution, where high probability is assigned to pixels that affect the composition of the image with respect to the human visual system. We formulate the crafting of adversarial examples as a two-player turn-based stochastic game, where the first player’s objective is to minimise the distance to an adversarial example by manipulating the features, and the second player can be cooperative, adversarial, or random. We show that, theoretically, the two-player game can converge to the optimal strategy, and that the optimal strategy represents a globally minimal adversarial image. For Lipschitz networks, we also identify conditions that provide safety guarantees that no adversarial examples exist. Using Monte Carlo tree search we gradually explore the game state space to search for adversarial examples. Our experiments show that, despite the black-box setting, manipulations guided by a perception-based saliency distribution are competitive with state-of-the-art methods that rely on white-box saliency matrices or sophisticated optimization procedures. Finally, we show how our method can be used to evaluate robustness of neural networks in safety-critical applications such as traffic sign recognition in self-driving cars.