Thomas Brunner

Tobias Kessler, Julian Bernhard, Martin Buechel et al.

Although many research vehicle platforms for autonomous driving have been built in the past, hardware design, source code and lessons learned have not been made available for the next generation of demonstrators. This raises the efforts for the research community to contribute results based on real-world evaluations as engineering knowledge of building and maintaining a research vehicle is lost. In this paper, we deliver an analysis of our approach to transferring an open source driving stack to a research vehicle. We put the hardware and software setup in context to other demonstrators and explain the criteria that led to our chosen hardware and software design. Specifically, we discuss the mapping of the Apollo driving stack to the system layout of our research vehicle, fortuna, including communication with the actuators by a controller running on a real-time hardware platform and the integration of the sensor setup. With our collection of the lessons learned, we encourage a faster setup of such systems by other research groups in the future.

Francesco Croce, Sven Gowal, Thomas Brunner et al.

Adaptive defenses, which optimize at test time, promise to improve adversarial robustness. We categorize such adaptive test-time defenses, explain their potential benefits and drawbacks, and evaluate a representative variety of the latest adaptive defenses for image classification. Unfortunately, none significantly improve upon static defenses when subjected to our careful case study evaluation. Some even weaken the underlying static model while simultaneously increasing inference computation. While these results are disappointing, we still believe that adaptive test-time defenses are a promising avenue of research and, as such, we provide recommendations for their thorough evaluation. We extend the checklist of Carlini et al. (2019) by providing concrete steps specific to adaptive defenses.

Thomas Brunner, Frederik Diehl, Alois Knoll

Many optimization methods for generating black-box adversarial examples have been proposed, but the aspect of initializing said optimizers has not been considered in much detail. We show that the choice of starting points is indeed crucial, and that the performance of state-of-the-art attacks depends on it. First, we discuss desirable properties of starting points for attacking image classifiers, and how they can be chosen to increase query efficiency. Notably, we find that simply copying small patches from other images is a valid strategy. We then present an evaluation on ImageNet that clearly demonstrates the effectiveness of this method: Our initialization scheme reduces the number of queries required for a state-of-the-art Boundary Attack by 81%, significantly outperforming previous results reported for targeted black-box adversarial examples.

Thomas Brunner, Frederik Diehl, Michael Truong Le et al.

Semantic Embeddings are a popular way to represent knowledge in the field of zero-shot learning. We observe their interpretability and discuss their potential utility in a safety-critical context. Concretely, we propose to use them to add introspection and error detection capabilities to neural network classifiers. First, we show how to create embeddings from symbolic domain knowledge. We discuss how to use them for interpreting mispredictions and propose a simple error detection scheme. We then introduce the concept of semantic distance: a real-valued score that measures confidence in the semantic space. We evaluate this score on a traffic sign classifier and find that it achieves near state-of-the-art performance, while being significantly faster to compute than other confidence scores. Our approach requires no changes to the original network and is thus applicable to any task for which domain knowledge is available.

Chih-Hong Cheng, Chung-Hao Huang, Thomas Brunner et al.

We study the problem of safety verification of direct perception neural networks, where camera images are used as inputs to produce high-level features for autonomous vehicles to make control decisions. Formal verification of direct perception neural networks is extremely challenging, as it is difficult to formulate the specification that requires characterizing input as constraints, while the number of neurons in such a network can reach millions. We approach the specification problem by learning an input property characterizer which carefully extends a direct perception neural network at close-to-output layers, and address the scalability problem by a novel assume-guarantee based verification approach. The presented workflow is used to understand a direct perception neural network (developed by Audi) which computes the next waypoint and orientation for autonomous vehicles to follow.

Frederik Diehl, Thomas Brunner, Michael Truong Le et al.

By interpreting a traffic scene as a graph of interacting vehicles, we gain a flexible abstract representation which allows us to apply Graph Neural Network (GNN) models for traffic prediction. These naturally take interaction between traffic participants into account while being computationally efficient and providing large model capacity. We evaluate two state-of-the art GNN architectures and introduce several adaptations for our specific scenario. We show that prediction error in scenarios with much interaction decreases by 30% compared to a model that does not take interactions into account. This suggests that interaction is important, and shows that we can model it using graphs. This makes GNNs a worthwhile addition to traffic prediction systems.

Thomas Brunner, Frederik Diehl, Michael Truong Le et al.

We consider adversarial examples for image classification in the black-box decision-based setting. Here, an attacker cannot access confidence scores, but only the final label. Most attacks for this scenario are either unreliable or inefficient. Focusing on the latter, we show that a specific class of attacks, Boundary Attacks, can be reinterpreted as a biased sampling framework that gains efficiency from domain knowledge. We identify three such biases, image frequency, regional masks and surrogate gradients, and evaluate their performance against an ImageNet classifier. We show that the combination of these biases outperforms the state of the art by a wide margin. We also showcase an efficient way to attack the Google Cloud Vision API, where we craft convincing perturbations with just a few hundred queries. Finally, the methods we propose have also been found to work very well against strong defenses: Our targeted attack won second place in the NeurIPS 2018 Adversarial Vision Challenge.