Georg Weißenbacher

Anagha Athavale, Samuel Teuber, Matteo Maffei et al.

The robustness of deep neural networks (DNNs) is critical in security-sensitive applications, where small input perturbations should not alter model predictions. This property is commonly formalized as local or global robustness: the former considers perturbations around a single input, while the latter -- strictly stronger -- quantifies over all input pairs. While local robustness can be expressed as a safety property, global robustness is a 2-safety property, making it substantially more challenging to verify. We present a novel static analysis technique for verifying the global robustness of DNNs. Our approach is based on differential halo zonotopes, a new abstract domain that extends zonotopes to jointly propagate pairs of perturbed inputs in lock-step while tightly bounding their divergence. In addition, we introduce a symmetric variant of confidence-based global robustness that disregards perturbations leading to differing but low-confidence predictions. This relaxation yields a practically meaningful notion of robustness that applies to a broader class of networks. We implement our approach in a new tool, called TwoSafe, and evaluate it on standard DNN verification benchmarks, including widely deployed models. Our results show that TwoSafe significantly outperforms the state of the art in both precision and scalability, enabling the verification of networks an order of magnitude larger than those handled by prior techniques.

Andreas Fellner, Mitra Tabaei Befrouei, Georg Weissenbacher

We present a new method for model-based mutation-driven test case generation. Mutants are generated by making small syntactical modifications to the model or source code of the system under test. A test case kills a mutant if the behavior of the mutant deviates from the original system when running the test. In this work, we use hyperproperties-which allow to express relations between multiple executions-to formalize different notions of killing for both deterministic as well as non-deterministic models. The resulting hyperproperties are universal in the sense that they apply to arbitrary reactive models and mutants. Moreover, an off-the-shelf model checking tool for hyperproperties can be used to generate test cases. We evaluate our approach on a number of models expressed in two different modeling languages by generating tests using a state-of-the-art mutation testing tool.

Andreas Holzer, Daniel Schwartz-Narbonne, Mitra Tabaei Befrouei et al.

Error invariants are assertions that over-approximate the reachable program states at a given position in an error trace while only capturing states that will still lead to failure if execution of the trace is continued from that position. Such assertions reflect the effect of statements that are involved in the root cause of an error and its propagation, enabling slicing of statements that do not contribute to the error. Previous work on error invariants focused on sequential programs. We generalize error invariants to concurrent traces by augmenting them with additional information about hazards such as write-after-write events, which are often involved in race conditions and atomicity violations. By providing the option to include varying levels of details in error invariants-such as hazards and branching information-our approach allows the programmer to systematically analyze individual aspects of an error trace.We have implemented a hazard-sensitive slicing tool for concurrent traces based on error invariants and evaluated it on benchmarks covering a broad range of real-world concurrency bugs. Hazard-sensitive slicing significantly reduced the length of the considered traces and still maintained the root causes of the concurrency bugs.