DeepGlobal: A framework for global robustness verification of feedforward neural networks

Feed forward neural networks (FNNs) have been deployed in a variety of domains, though achieving great success, also pose severe safety and reliability concerns. Existing adversarial attack generation and automatic verification techniques cannot formally verify a network globally, i.e., finding all adversarial dangerous regions (ADRs) of a network is out of their reach. To address this problem, we develop a global robustness verifiable FNN framework DeepGlobal with four components: 1) a rule-generator finding all potential boundaries of a network by logical reasoning; 2) a new network architecture Sliding Door Network (SDN) enabling rule generation in a feasible way; 3) a selector which selects real boundaries from the generated potential boundaries; 4) a filter finding ADRs with meaningful adversarial examples. The ADRs can be represented by the identified real boundaries. We demonstrate the effectiveness of our approach on both synthetic and real datasets.