Latency-sensitive edge AI services must balance strict deadlines, output quality, and limited compute and energy budgets. However, static CPU provisioning wastes resources because inference cost varies substantially across inputs, model variants, and runtime conditions. We present PRISM, a prediction-guided runtime framework that jointly selects model variants and CPU allocations for containerized edge microservices. Using container-level energy monitoring and lightweight regression models, PRISM adapts each pipeline stage in place and minimizes predicted CPU-package energy under deadline, resource, and offline model-level Quality of Result (QoR) constraints. We evaluate PRISM on more than 52,000 requests in an Automatic License Plate Recognition (ALPR) pipeline with detection and recognition stages. For detection, PRISM reduces energy consumption by 36

Edge nodes are capacity-constrained by design, yet many edge workloads can trade result quality for resource efficiency at runtime. Existing edge pricing mechanisms largely treat requests as fixed-configuration submissions and rarely exploit per-request quality flexibility under overload. We present PRICE, an incentive mechanism that couples a utilization-dependent price signal to per-request quality selection. As utilization increases, rising acceptance prices make resource-intensive variants less likely to be selected, shifting accepted requests toward lighter execution and allowing the node to serve significantly more requests while operating near capacity. Evaluation on real hardware under sustained overload shows that PRICE outperforms both fixed-allocation and dynamic-pricing baselines in accepted throughput and CPU utilization. The results are robust across pricing function families, task-duration distributions, and client populations. Result-quality flexibility is a powerful but underused control dimension for overload management at the edge, and pricing is an effective mechanism to exploit it.

Training machine learning models for city-wide mobility predictions typically requires extensive local trajectory data for training, or transfer learning techniques to adapt from data-rich to data-scarce cities. The resulting models are tailored to the specific target city for which at least some available trajectory data is necessary, which makes generalization to cities without any mobility data largely infeasible. In prior work, we proposed H0xtra, a hypernetwork-based model architecture that enables zero-shot transfer to unseen cities without requiring any target city trajectory data. In this work, we introduce HyperAOE, a generalization of H0xtra, provide an in-depth analysis of its capabilities, and demonstrate its applicability across a broad range of settings. Specifically, HyperAOE decomposes H0xtra into its three general components: (i) spatial meta data, such as points of interest information or population density data; (ii) a city encoder that transforms this spatial information into location embeddings; and (iii) a sequence-to-sequence backbone that captures sequential patterns in human trajectories, and formalizes their combination in a unified framework. This general setup enables the use of diverse types of spatial meta data, various city encoder architectures, and a wide range of sequence backbone architectures. Moreover, it allows for the plug-in extension of existing urban trajectory modeling approaches, equipping them with zero-shot cross-city transfer capabilities off the shelf. Through extensive experiments, we analyze the transfer capabilities and limitations of HyperAOE. Our results demonstrate its effectiveness across a variety of experimental settings and highlight its strong zero-shot transfer performance. Our code can be accessed at https://github.com/DLR-Prot-of-Terrestrial-Infrastructures/HyperAOE.

System auditing on Android faces two problems. First, existing syscall tracers lose events under load, silently overwriting entries faster than a user space reader can drain them. Second, security-relevant application behavior is mediated through Binder, Android's kernel IPC mechanism, and is therefore hidden from the syscall layer. The Binder parcels that the kernel does see carry no method names or typed arguments, a disconnect between low-level events and high-level behavior known as the semantic gap. Existing approaches address the semantic gap either by modifying the Android platform, making them difficult to adjust to OS updates, or by instrumenting the traced application in user space, which sophisticated adversaries can evade by bypassing the instrumented framework APIs. We present WOOTdroid, a design and prototype for on-device tracing on stock Android that addresses both problems without OS modification or application instrumentation. WDSys, an eBPF port of eAudit-style syscall auditing, runs on current Android with at most 3.6

AI-assisted decision-making is subject to AI model uncertainty. Prior works proposed to make this uncertainty explicit for increasing trust and transparency, but its behavioral role was rarely treated. This position paper argues, from a game-theoretic perspective, that human–AI decision support should be viewed as a repeated mechanism in which AI uncertainty functions as a strategic signal that shapes how users adopt reliance policies over time. We formalize a framework in which the interface specifies uncertainty signals, user response such as accepting versus verifying, and the resulting policy-shaping consequences. These repeated steps are used to characterize near-separating reliance regimes. A first pilot study conducted with 180 participants supports our proposition: Our game-theoretic mechanism increased verification and sharply reduced blind acceptance of wrong AI outputs. These initial results support treating human–AI interaction as a game-theoretic mechanism with uncertainty as a strategic signal, rather than a static model property or purely informational label.

This paper contributes an in-depth investigation (N=24) of privacy perceptions in the context of medical data donation apps. Medical data donation refers to the act of voluntarily sharing medical data with research institutions, which plays a crucial role in advancing healthcare research and personalized medicine. To design effective medical data donation apps, we need to understand how privacy expectations affect people's willingness to use such apps. We focus on non-users—those who have no experience with medical data donation apps—because gaining a deeper understanding of their perceptions is essential for fostering the adoption of these apps. Our findings highlight the importance of trust, transparency, and anonymity as driving factors. Participants expressed a willingness to share highly sensitive medical data with the apps if they were assured of complete anonymity, yet criticism regarding the risks of de-anonymization was also raised. Based on our results, we identify privacy awareness issues, especially concerning data sensitivity. Additionally, we explain the differences between participants' privacy expectations and preferences and what existing medical data donation apps offer. Finally, we provide guidance for the development of future user-centric medical data donation apps.

Large Language Models (LLMs) show growing promise in autonomous driving by reasoning over complex traffic scenarios to generate path plans. However, their tendencies toward overconfidence, and hallucinations raise critical safety concerns. We introduce SafePath, a modular framework that augments LLM-based path planning with formal safety guarantees using conformal prediction. SafePath operates in three stages. In the first stage, we use an LLM that generates a set of diverse candidate paths, exploring possible trajectories based on agent behaviors and environmental cues. In the second stage, SafePath filters out high-risk trajectories while guaranteeing that at least one safe option is included with a user-defined probability, through a multiple-choice question-answering formulation that integrates conformal prediction. In the final stage, our approach selects the path with the lowest expected collision risk when uncertainty is low or delegates control to a human when uncertainty is high. We theoretically prove that SafePath guarantees a safe trajectory with a user-defined probability, and we show how its human delegation rate can be tuned to balance autonomy and safety. Extensive experiments on nuScenes and Highway-env show that SafePath reduces planning uncertainty by 77% and collision rates by up to 70%, demonstrating effectiveness in making LLM-driven path planning more safer.

Understanding and predicting human mobility during disasters is crucial for effective disaster management. Knowledge about population locations can greatly enhance rescue missions and evacuations. Realistic models that reflect observable mobility patterns and volumes are crucial for estimating population locations. However, existing models are limited in their applicability to disasters, as they are typically restricted to describing regular mobility patterns. Machine learning models trained to capture patterns observable in provided training data also face this limitation. The necessity of large amounts of training data for machine learning models, coupled with the scarcity of data on mobility in disasters, often constrains the feasibility of their training. Various strategies have been developed to overcome this issue, which we present and discuss in this systematic literature review. Our review aims to support and accelerate the synthesis of novel approaches by establishing a knowledge base for future research. This review identified a condensed field of related contributions exhibiting high methodology and context diversity. We classified and analyzed the relevant contributions based on their proposed approach and subsequently discussed and compared them qualitatively. Finally, we elaborated on general challenges and highlighted areas for future research.

Realistic crowd simulations are essential for immersive virtual environments, relying on both individual behaviors (microscopic dynamics) and overall crowd patterns (macroscopic characteristics). While recent data-driven methods like deep reinforcement learning improve microscopic realism, they often overlook critical macroscopic features such as crowd density and flow, which are governed by spatio-temporal spawn dynamics, namely, when and where agents enter a scene. Traditional methods, like random spawn rates, stochastic processes, or fixed schedules, are not guaranteed to capture the underlying complexity or lack diversity and realism. To address this issue, we propose a novel approach called nTPP-GMM that models spatio-temporal spawn dynamics using Neural Temporal Point Processes (nTPPs) that are coupled with a spawn-conditional Gaussian Mixture Model (GMM) for agent spawn and goal positions. We evaluate our approach by orchestrating crowd simulations of three diverse real-world datasets with nTPP-GMM. Our experiments demonstrate the orchestration with nTPP-GMM leads to realistic simulations that reflect real-world crowd scenarios and allow crowd analysis.

Research in Augmented Reality (AR) and Virtual Reality (VR) has mostly viewed them in isolation. Yet, when used together in practical settings, AR and VR each offer unique strengths, necessitating multiple transitions to harness their advantages. This paper investigates potential challenges in Cross-Reality (CR) transitions to inform future application design. We implemented a CR system featuring a 3D modeling task that requires users to switch between PC, AR, and VR. Using a talk-aloud study (n=12) and thematic analysis, we revealed that frictions primarily arose when transitions conflicted with users’ Spatial Mental Model (SMM). Furthermore, we found five transition archetypes employed to enhance productivity once an SMM was established. Our findings uncover that transitions have to focus on establishing and upholding the SMM of users across realities, by communicating differences between them.

Mobile and IoT applications increasingly adopt deep learning inference to provide intelligence. Inference requests are typically sent to a cloud infrastructure over a wireless network that is highly variable, leading to the challenge of dynamic Service Level Objectives (SLOs) at the request level.

Market forces such as rising amounts of product variants and decreasing batch sizes lead to higher complexity in manufacturing processes. Therefore, production management's demand for data-based process transparency is growing continuously as well as the number of companies turning to process mining to address these challenges. Information systems in production usually do not provide readily available event log data for the analysis. This paper investigates several techniques for inferring missing event log data in production processes by extracting events with timestamps from sensor data from machines and link them to process instances. We demonstrate the effectiveness of our approach in a real-world manufacturing environment. The evaluation of the resulting event logs revealed that the quality of the timestamps and the assignment of the actual process instances is sufficient to apply process mining techniques that would have required both greater effort and higher cost intensity if a traceability system had been implemented.

Medical data donation involves voluntarily sharing medical data with research institutions, which is crucial for advancing healthcare research. However, the sensitive nature of medical data poses privacy and security challenges. The primary concern is the risk of de-anonymization, where users can be linked to their donated data through background knowledge or communication metadata. In this paper, we introduce Anonify, a decentralized anonymity protocol offering strong user protection during data donation without reliance on a single entity. It achieves dual-level anonymity protection, covering both communication and data aspects by leveraging Distributed Point Functions, and incorporating k-anonymity and stratified sampling within a secret-sharing-based setting. Anonify ensures that the donated data is in a form that affords flexibility for researchers in their analyses. Our evaluation demonstrates the efficiency of Anonify in preserving privacy and optimizing data utility. Furthermore, the performance of machine learning algorithms on the anonymized datasets generated by the protocol shows high accuracy and precision.

Personalized privacy assistants (PPAs) communicate privacy-related decisions of their users to Internet of Things (IoT) devices. There are different ways to implement PPAs by varying the degree of autonomy or decision model. This paper investigates user perceptions of PPA autonomy models and privacy profiles – archetypes of individual privacy needs – as a basis for PPA decisions in private environments (e.g., a friend’s home). We first explore how privacy profiles can be assigned to users and propose an assignment method. Next, we investigate user perceptions in 18 usage scenarios with varying contexts, data types and number of decisions in a study with 1126 participants. We found considerable differences between the profiles in settings with few decisions. If the number of decisions gets high (> 1/h), participants exclusively preferred fully autonomous PPAs. Finally, we discuss implications and recommendations for designing scalable PPAs that serve as privacy interfaces for future IoT devices.

Advancements in haptics for Virtual Reality (VR) increased the quality of immersive content. Particularly, recent efforts to provide realistic temperature sensations have gained traction, but most often require very specialized or large complex devices to create precise thermal actuations. However, being largely detached from the real world, such a precise correspondence between the physical location of thermal stimuli and the shown visuals in VR might not be necessary for an authentic experience. In this work, we contribute the findings of a controlled experiment with 20 participants, investigating the spatial localization accuracy of thermal stimuli while having matching and non-matching visual cues of a virtual heat source in VR. Although participants were highly confident in their localization decisions, their ability to accurately pinpoint thermal stimuli was notably deficient.