Abstract Threat Modeling is an essential step in secure software system development. It is a (so far) manual, attacker-centric approach for identifying architecture-level security flaws during the planning phase of software systems. In recent years, academia has presented ideas to automate threat detection that do not focus on a particular class of security flaws but offer means of pattern-based security flaw descriptions. However, comparing presented ideas (tools) for automated threat detection contains the potential for unwilling bias or restricted information content. In this work, we investigate the process of comparing automatic security flaw detection tools, clarify common pitfalls during this process, and propose a fair, reproducible, and informative comparison approach to be used as a community standard. We additionally discuss the necessary steps for the community to effectively implement this approach and support improved comparisons and evaluations in the future. We use a previously published case study to determine problems with current comparison techniques and classify different levels of comparison to be used for future reference as our main contribution. As a consequence, we propose using a model-based approach for specifying security flaws and apply an existing natural language-based catalogue to this model-based approach. Furthermore, we introduce an inspection process model (for providing a standard to specify findings of a threat detection process) to streamline the evaluation and comparisons of automatic security flaw detection tools. We provide an exemplary evaluation of this detection guideline and inspection process model along the lines of both automatic approaches from the original case study. All artefacts of the work are publicly available to support the research community and to create a common baseline for future tool comparisons.

Trust in electronic devices is dependent on their safe and reliable behavior. An integral part is the correct design of the hardware. While classically simulation-based approaches have been applied, only through formal proof techniques complete correctness can be guaranteed. The core of these formal approaches, and responsible for time and space complexity, is the choice of the underlying data structure to represent the functional behavior. A significant class of data structures are graph-based function representations, like BDDs, KFDDs or *BMDs. These have shown excellent properties – provability in polynomial time and space – for some function classes, e.g., adders. Experimental studies have validated these properties, and formal proofs can guarantee this behavior. Unfortunately, these properties often cannot be generalized to varying function classes. One reason is that graph-based representations are usually tailored for either bit-level or word-level functions. However, designing hybrid data structures that can represent both types in parallel might allow formal proofs for even larger functional classes.In this paper, we demonstrate how to design these hybrid data structures, overcoming limitations of current formal verification approaches. We introduce a generalized concept on decompositions and graph-based function representations based on Kronecker matrices with an extended element space and dimension. It is shown how these extensions allow the representation of hybrid function classes, paving the way for more trust in electronic devices.

Magnetic Resonance Imaging (MRI) is an essential tool for medical diagnosis. At the same time, its usage requires profound expert knowledge to determine the ideal MR sequence and protocol to be run. Until now, the contrast and quality of the resulting image have relied mainly on the radiologist's expertise. When confronted with clinical requirements and patient information, the radiologist chooses suitable sequence protocols for the examination. We propose a workflow that supports medical personnel in finding the optimal sequence for a given diagnostic task. To that end, we combine evolutionary algorithms for the optimisation, machine learning techniques for training a surrogate optimisation function from simulated MRI data, and domain-specific languages to allow non-programmers to formulate their requirements and constraints semi-formally. In this paper, we focus on the efficient usage of real-world application-motivated adaptions of the used evolutionary algorithm and evaluate their effects on four real-life sequence examples. We show that it is essential to use an adaption for the surrogate model to obtain realistic solutions and use correlation information about the search space to stay in feasible areas of the search space and thus improve optimisation quality. These findings are a first step in automating the entire MRI-sequence optimisation flow, which is necessary to allow a more widespread usage of this essential medical diagnostic technique.

Adapting optimisation algorithms, such as evolutionary algorithms, to a problem is a necessity. The required collection and exchange of domain information is an important but tedious task in real-world projects involving several experts from different areas of expertise (e.g. the domain and the optimisation area). This paper presents a structured approach that allows the experts to systematically provide their knowledge using domain-specific languages. The presented approach defines a different language for the involved experts that enables them to add their knowledge and use the information provided by other experts. The languages are extensible, allowing the addition of new optimisation aspects without changing the actual language. These languages are the front-end to a versatile open-source optimisation tool, that we built, enabling the actual execution of the optimisation. It additionally provides features for surrogate models as well as data generation and different benchmarks for evaluation. Conducting a user study, we show that the language is suitable to express the domain knowledge and domain experts can use the language to describe their domain knowledge after a short introduction. This way, the approach reduces the effort for domain experts in providing their information. As a side effect, the complete configuration of the optimisation execution through these languages allows an easy and reliable reproduction.

General-Purpose Computation on Graphics Processing Units (GPGPUs) are becoming crucial in accelerating computing capacity. Due to the massive parallelism capabilities of GPUs, they can achieve impressive speedups of up to 32 times compared to common CPUs. However, writing highly parallel code and utilizing a GPU is challenging for programmers. Developers are facing new challenges since GPUs handle threads and parallelism differently from CPUs. Academia and industry proposed several profilers to support developers in terms of code optimization. These profilers often require an actual device (e.g., GPU) and take a long time for the profiling process. We propose HyPA, a hybrid Parallel Thread Execution (PTX) Analyzer that inspects PTX code statically and dynamically. HyPA implements a partly functional emulator that executes instructions that rely on runtime dependencies to count the number of executed PTX instructions and divergent branches. HyPa executes compiled kernels—the programs that run on GPUs—generated by the CUDA compiler and supports the full PTX 7.7 specification. Our functional emulator allows significantly faster analysis of PTX code compared to standard profilers. In our evaluation, we quantify this increase in performance through benchmark runs. HyPA achieved speedups of up to 536% compared to the nvprof profiler. Moreover, our approach can gather performance metrics beyond static analysis (e.g., branch efficiency) by a faster execution time than by profiling the application on an actual device. Finally, we provide an open-source implementation of HyPA to help developers and system designers in further research and development.

Evolutionary algorithms are a well-known optimisation technique, especially for non-convex, multi-modal optimisation problems. Their capability of adjusting to different search spaces and tasks by choosing the suitable encoding and operators has led to their widespread use in various application domains. However, application domains sometimes come with difficulties like fitness functions that can not be evaluated or not more than a few times. In these situations, surrogate functions or approximative fitness functions allow the evolutionary algorithm to work despite this complication. Still, using approximative fitness functions comes with a price: The fitness value is no longer correct for every individual, and the algorithm can not know which value to trust. However, statistical methods yield knowledge about the preciseness of the approximation. We propose using this knowledge to adapt the fitness value to ease the effects of the approximative nature. We choose to use the information given in the density of the training data, which has computational merits over the use of other techniques like cross-validation or prediction intervals. We evaluate our method on four well-known benchmark functions and achieve good optimisation success and computation time results.

Evolutionary algorithms are a successful application of bio-inspired behaviour in the field of Artificial Intelligence. Transferring mechanisms such as selection, mutation, and recombination, evolutionary algorithms are capable of surmounting the disadvantages of traditional methods. Adjusting an evolutionary algorithm to a specific problem requires both, a good understanding of the problem and deep knowledge of the effects of choosing one or another operator in the algorithm. This becomes an especially difficult task when the fitness function is not analytically given - that is, exists only as an approximation, that is highly dependent on the present training data. We propose using prediction intervals to modify the fitness function such, that worse fitness values are less penalized if they occur in a poorly fitted area. We evaluate this with an example from material sciences as well as four standard benchmark algorithms for evolutionary algorithms using a Support Vector Regression for training the approximative fitness function and find that our approach outperforms the naive approximative function.

Evolutionary algorithms are a very general method for optimization problems that allow adaption to many different use cases. Application to real-world problems usually comes with features as constraints, dependencies and approximations. When a multidimensional search space comes with strings attached- namely dependencies between its dimensions- an expression in two ways is possible: Restrictive-as equalities or inequalities- or vague-as correlations between dimensions, for example. Correlations between dimensions are not as easy to grasp as constraints. Therefore, well-known techniques as death penalty or penalty functions do not apply directly. We propose new mutation and recombination operators that incorporate domain knowledge to increase the offspring fraction that adheres to these correlations. We evaluate our approach with several benchmark functions and different assumptions on the dependencies of the search space. We compare the likelihood of valid (in terms of adhering correlations) outcomes of algorithms using standard mutation and recombination operators to those with the proposed operators. We find that the correlation-aware operators preserve population's features in terms of dependencies.

The massive increase in computation power leads to a renaissance of supervised learning techniques, which were published decades ago but have so far been confined to theory. These techniques form the increasingly important field of Machine Learning (ML), which contributes to a large variety of research concerning industrial, automotive but also consumer applications strongly influencing our daily life. Commonly, the learning techniques require a set of labeled data, which involves a resource-intensive generation, to conduct the training. Depending on the dimensionality of the data and the required precision as needed by the application, the amount of training data varies. In case of insufficient training data, the prediction is of low-quality or not even possible at all, restricting the applicability of ML. This work proposes a combination of formal techniques and ML to implement a framework that allows coping with high-dimensional, training data while retaining a high prediction quality. The efficacy of this method is exemplarily demonstrated on the basis of an interdisciplinary material science research problem concerning the development of new structural materials, though it can be adapted to further applications.

Cryptographic devices that potentially operate in hostile physical environments need to be secured against side-channel attacks. In order to ensure the effectiveness of the required countermeasures, scientists, developers, and evaluators need efficient methods to test the security level of a device. In this paper we propose a new framework based on confidence intervals that extends established t-test based approaches for test-vector leakage assessment (TVLA). In comparison to previous TVLA approaches the new methodology does not only enable the detection of leakage but can also assert its absence. The framework is robust against noise in the evaluation system and thereby avoids false negatives. These improvements can be achieved without overhead in measurement complexity and with a minimum of additional computational costs compared to previous approaches. We evaluate our method under realistic conditions by applying it to a protected implementation of AES.