Adversarial attacks against deep neural networks are commonly constructed under ℓ_p norm constraints, most often using p=1, p=2 or p=∞, and potentially regularized for specific demands such as sparsity or smoothness. These choices are typically made without a systematic investigation of how the norm parameter p influences the structural and perceptual properties of adversarial perturbations. In this work, we study how the choice of p affects sparsity and smoothness of adversarial attacks generated under ℓ_p norm constraints for values of p ∈ [1,2]. To enable a quantitative analysis, we adopt two established sparsity measures from the literature and introduce three smoothness measures. In particular, we propose a general framework for deriving smoothness measures based on smoothing operations and additionally introduce a smoothness measure based on first-order Taylor approximations. Using these measures, we conduct a comprehensive empirical evaluation across multiple real-world image datasets and a diverse set of model architectures, including both convolutional and transformer-based networks. We show that the choice of ℓ_1 or ℓ_2 is suboptimal in most cases and the optimal p value is dependent on the specific task. In our experiments, using ℓ_p norms with p∈ [1.3, 1.5] yields the best trade-off between sparse and smooth attacks. These findings highlight the importance of principled norm selection when designing and evaluating adversarial attacks.

The ability to explain the outputs of deep neural networks is crucial, particularly in safety-critical applications. Recently, complex-valued neural networks (CVNNs) have gained increasing popularity; however no dedicated framework for explaining their predictions has been introduced. To address this gap, we develop DeepCSHAP, a method that extends the widely used SHAP framework to CVNNs. In addition, we adapt common gradient-based explanation techniques to the complex domain using Wirtinger derivatives and provide both these methods and DeepCSHAP in an open-source explainability library. Experimental results demonstrate that DeepCSHAP outperforms other explanation methods in explaining model outputs of CVNNs.

Purpose Before selective internal radiation therapy (SIRT), 99mTc macroaggregated albumin (MAA) particles are injected from the same catheter position(s) as a surrogate for later resin sphere distribution and to enable predictive dosimetry. Deviations in tumor segmentation affect predicted tumor and normal-liver absorbed doses and therefore prescribed activity. This study investigates the magnitude of interobserver variability in tumor segmentation for inexperienced and experienced observers when using biphasic contrast-enhanced CT as part of the 99mTc-MAA SPECT/CT imaging protocol and how it impacts the resulting tumor and normal-liver doses. Methods One inexperienced observer (performing two segmentations eight weeks apart) and one experienced observer used MIM SurePlan LiverY90 software to create tumor regions of interest (ROIs) from the SPECT/CT data of 19 patients. These three sets of ROIs were compared to the original clinical ROIs that had been defined similarly by another experienced physician together with a physicist. The Dice Similarity Coefficient (DSC) was calculated as a measure of tumor ROI overlap. Additionally, the resulting tumor and normal-liver dose differences using the partition model (assumed homogeneity within tumor and normal-liver compartments, doses represented by their mean values) between the three retrospectively determined sets of ROIs and the clinical ROIs were calculated and compared between inexperienced and experienced observers. Results DSC values between 0.73 and 0.75 were observed, indicating moderate to good agreement of segmentations, for inexperienced and experienced observers alike. Tumor dose differences were -3±13%, 2±10%, 2±10%; normal-liver dose differences were 9±16%, -9±14%, -4±14%, showing no relevant differences between inexperienced and experienced observers. Conclusion Significant interindividual variability in tumor segmentation and thus non-negligible deviations in tumor and normal-liver doses exist even with high-quality SPECT/CT imaging; however, for inexperienced and experienced observers alike. Hence, inexperienced observers with limited training can perform acceptably well. The magnitude of this variability should be considered when choosing injected activity based on predictive dosimetry, nevertheless.

Complex-valued neural networks (CVNNs) are rising in popularity for all kinds of applications. To safely use CVNNs in practice, analyzing their robustness against outliers is crucial. One well known technique to understand the behavior of deep neural networks is to investigate their behavior under adversarial attacks, which can be seen as worst case minimal perturbations. We design Phase Attacks, a kind of attack specifically targeting the phase information of complex-valued inputs. Additionally, we derive complex-valued versions of commonly used adversarial attacks. We show that in some scenarios CVNNs are more robust than RVNNs and that both are very susceptible to phase changes with the Phase Attacks decreasing the model performance more, than equally strong regular attacks, which can attack both phase and magnitude.

Complex-Valued Neural Networks are a rising field in research and applications. However, the lack of pretrained complex-valued models complicates their training process, since users can not benefit from the advantages of having large pretrained models publicly available. We present three methods to utilize readily available pretrained real-valued models to initialize complex-valued models. We show that this consistently outperforms training the Complex-Valued Neural Networks from scratch. Additionally, we show that in some scenarios Complex-Valued Neural Networks can outperform their real-valued counterparts, confirming their suitability as an alternative for various applications.

In this work we propose CVKAN, a complex-valued Kolmogorov-Arnold Network (KAN), to join the intrinsic interpretability of KANs and the advantages of Complex-Valued Neural Networks (CVNNs). We show how to transfer a KAN and the necessary associated mechanisms into the complex domain. To confirm that CVKAN meets expectations we conduct experiments on symbolic complex-valued function fitting and physically meaningful formulae as well as on a more realistic dataset from knot theory. Our proposed CVKAN is more stable and performs on par or better than real-valued KANs while requiring less parameters and a shallower network architecture, making it more explainable.

Deep Neural Networks are widely used in academy as well as corporate and public applications, including safety critical applications such as health care and autonomous driving. The ability to explain their output is critical for safety reasons as well as acceptance among applicants. A multitude of methods have been proposed to explain real-valued neural networks. Recently, complex-valued neural networks have emerged as a new class of neural networks dealing with complex-valued input data without the necessity of projecting them onto $\mathbb{R}^2$. This brings up the need to develop explanation algorithms for this kind of neural networks. In this paper we provide these developments. While we focus on adapting the widely used DeepSHAP algorithm to the complex domain, we also present versions of four gradient based explanation methods suitable for use in complex-valued neural networks. We evaluate the explanation quality of all presented algorithms and provide all of them as an open source library adaptable to most recent complex-valued neural network architectures.

Most deep learning pipelines are built on real-valued operations to deal with real-valued inputs such as images, speech or music signals. However, a lot of applications naturally make use of complex-valued signals or images, such as MRI or remote sensing. Additionally the Fourier transform of signals is complex-valued and has numerous applications. We aim to make deep learning directly applicable to these complex-valued signals without using projections into ℝ^2. Thus we add to the recent developments of complex-valued neural networks by presenting building blocks to transfer the transformer architecture to the complex domain. We present multiple versions of a complex-valued Scaled Dot-Product Attention mechanism as well as a complex-valued layer normalization. We test on a classification and a sequence generation task on the MusicNet dataset and show improved robustness to overfitting while maintaining on-par performance when compared to the real-valued transformer architecture.

One well established method of interactive image segmentation is the random walker algorithm. Considerable research on this family of segmentation methods has been continuously conducted in recent years with numerous applications. These methods are common in using a simple Gaussian weight function which depends on a parameter that strongly influences the segmentation performance. In this work we propose a general framework of deriving weight functions based on probabilistic modeling. This framework can be concretized to cope with virtually any parametric noise model. It eliminates the critical parameter and thus avoids time-consuming parameter search. We derive the specific weight functions for common noise types and show their superior performance on synthetic data as well as different biomedical image data (MRI images from the NYU fastMRI dataset, larvae images acquired with the FIM technique). Our framework could also be used in multiple other applications, e.g., the graph cut algorithm and its extensions.

The random walker method for image segmentation is a popular tool for semi-automatic image segmentation, especially in the biomedical field. However, its linear asymptotic run time and memory requirements make application to 3D datasets of increasing sizes impractical. We propose a hierarchical framework that, to the best of our knowledge, is the first attempt to overcome these restrictions for the random walker algorithm and achieves sublinear run time and constant memory complexity. The goal of this framework is– rather than improving the segmentation quality compared to the baseline method– to make interactive segmentation on out-of-core datasets possible. The method is evaluated quantitatively on synthetic data and the CT-ORG dataset where the expected improvements in algorithm run time while maintaining high segmentation quality are confirmed. The incremental (i.e., interaction update) run time is demonstrated to be in seconds on a standard PC even for volumes of hundreds of gigabytes in size. In a small case study the applicability to large real world from current biomedical research is demonstrated. An implementation of the presented method is publicly available in version 5.2 of the widely used volume rendering and processing software Voreen (https://www.uni-muenster.de/Voreen/).

Existing online social networks hardly care about users’ privacy rights. In particular, they do not permit users to keep control over “their” data. By “their” data, we denote data that refers to the respective user as an identifiable object within (textual, audio, image or video) media. The well-known concept of “usage control” employs a usage rights’ perspective (e.g. DRM), but it does not explicitly deal with privacy. In this paper, we instead propose the concept of “data control”, which exactly focusses on privacy rights and therefore employs a control rights’ perspective. Based on data control, we propose a defamation-free network (DFN) in which control rights are not only manifest and visible, but can also be exercised. We examine the main usage scenarios of such a network, and discuss the possible approaches for implementing it. Finally, we sketch a solution with an underlying P2P architecture and highlight the basic technological challenges and requirements.

In everyday life, trust is largely built from experience. Reputation-based trust models have been developed to formalize this concept. The application to networks like the Internet where a very large number of predominantly unknown principal identities engage in interactions is appealing considering that the evaluation of trusted experience may result in a more successful choice of trusted parties to interact with. In this paper we pick the SECURE framework, as developed within the equally named EU project on Global Computing, which builds upon event structures to model possible outcomes of interactions. We extend it by three concepts: (i) a flexible way to determine a degree of trust from given past behavior, (ii) a basic notion of context, exemplarily in the form of roles the interacting parties may occupy, and (iii) we explicitly equip observed events with a time component to refine the granularity of observations. We extend definitions of concepts used in SECURE in order to incorporate our notion of context information, we provide the syntax and semantics of an LTL-like logic, in its basics similar to the one proposed by Krukow, Nielsen and Sassone, that allows for layered reasoning about context information. We then show how this new language relates to the one used in SECURE and we determine under which conditions our concept of deriving trust from experience may be used within SECURE’s computational model to obtain a global state of trust.