Informatik
Refine
Document Type
- Article (36)
- Preprint (6)
- Conferenceobjectvolume (2)
- Master's Thesis (2)
- Bachelor Thesis (1)
- Bookcollection (1)
Has Fulltext
- yes (48)
Keywords
A photo lineup is an identification procedure widely used in criminal investigations. It involves presenting a suspect embedded into a set of known-innocent candidates (fillers) to an eyewitness in order to identify the suspect. Compiling fair lineups remains a challenge, particularly when the investigation team has to resort to unsuitable fillers. To support real-world lineups and mitigate the risk of misidentification, we present a practical approach for generating synthetic fillers. By injecting weighted layer-wise noise into a suspects latent vector representation, our approach generates visually distinct fillers while largely preserving demographic characteristics of the suspect. To assess suitability for investigative scenarios, we conducted a large-scale human perception study involving over 450 participants. The results show that the use of synthetically generated fillers leads to balanced identification performance, without making the suspect indistinguishable or stand out. Further experiments show that, compared to previous work, our approach achieves improved preservation of demographic characteristics. Overall, our work contributes to improving fairness in lineups and opens up avenues for supporting criminal investigations through the use of synthetic data.
Hash-functions are widely spread in computer science and used to map arbitrary large data to bit strings of a fixed length called a fingerprint. Cryptographic hash-functions like SHA1 or MD5 are established in various fields but have one ‘drawback’: if one bit of the input is changed, this result in a completely different fingerprint. As in computer science, investigators often look for similarities, using a ‘similarity function’ named fuzzy-hashing. The aim is to recognize the similarity of two files based on their fingerprint. For example, to make a statement like: both files have the same origin and have a similarity of 98%. In 2006, Jesse Kornblum presented an implementation for this idea. Since then today he has been quoted numerous times and his approach has been improved with respect to performance. However, there is no security analysis in terms of the reliability of the results. After analyzing his paper and the referenced implementation called ssdeep, we have discovered some vulnerabilities. As a start, the first two chapters will explain all basics about the paper and the implementation. Then some interesting characteristics of ssdeep will be discussed in chapter 3. The core of this paper is chapter 4 which introduces various methods to attack this approach. Finally, some ideas for improvements are suggested followed by a general consideration of Kornblum’s approach named context triggered piecewise hashing.
Biometrics has become a viable and popular solution for applications which require secure authentication. In spite of the advantages of biometrics as an automatic authentication technology, the usage of biometric characteristics raises significant concerns regarding personal data and privacy of subjects in these systems. To address these concerns, several biometric template protection schemes have been proposed in the literature to achieve trustworthy, reliable and privacy-preserving systems. In this paper, we benchmark several cancelable biometrics (CB) schemes on different biometric characteristics. We consider BioHashing, Multi-Layer Perceptron (MLP) hashing, Bloom filters, and two schemes based on Index-of-Maximum (IoM) hashing (i.e. IoM-URP and IoM-GRP). In addition to the mentioned CB schemes, we introduce a CB scheme (as a baseline) based on user-specific random transformations followed by binarization. We evaluate the unlinkability, irreversibility, and recognition performance (which are the required criteria by the ISO/IEC 24745 standard) of these CB schemes on deep learning-based templates extracted from different physiological and behavioural biometric characteristics including face, voice, finger vein, and iris. Our experiments show that all the studied CB schemes are almost unlinkable for different characteristics. We also observe that the mutual information (MI) between protected and unprotected templates varies according to the scenario and biometric characteristic. In terms of recognition accuracy, our study shows that deep templates protected by Bloom filters suffer from a drop in performance, while other CB schemes achieve competitive accuracies for different biometric characteristics. We provide an open-source implementation of all the experiments presented to facilitate the reproducibility of our results: https://github.com/otroshi/benchmark_cb .
Legacy systems are business-critical software products that are difficult to understand and modify. Many are written in dynamically typed languages like JavaScript or Python, which lack the benefits of static type annotations for documentation, error prevention, and maintainability. Manually retrofitting types to large legacy codebases requires significant effort, often taking years. Existing automated approaches have limited practical applicability: static type inference cannot handle dynamic language features, dynamic inference requires comprehensive test suites, and probabilistic inference is inherently unsound. No prior approach achieves both high type coverage and correctness at the repository level.
In this thesis, I investigate the use of agentic Artificial Intelligence (AI)—systems composed of Large Language Model (LLM)-based agents—to enable automated typing of legacy software projects. Through a systematic literature review, I identify the challenges and limitations of existing approaches. To overcome these, I propose a three-phase typing approach: (1) type-checking setup with initial error resolution, (2) type annotation and definition generation, and (3) strict type checking with refactoring. I implement the first two phases in AgenticTyper, an open-source tool for JavaScript and TypeScript repositories. My key innovation is a transpilation-based behavior preservation mechanism that prevents unintended modifications and addresses trust issues when using LLMs on production code.
I evaluate the approach through a case study on three proprietary legacy repositories with 456,000 lines of code. Phase one successfully sets up type checking and resolves all 633 initial errors for $25 in LLM API costs, compared to eight hours of manual effort for a single repository. Phase two achieves 86-96% type coverage with zero remaining type errors, costing $4 to $95 and requiring 4-70 hours of human review time depending on repository size. AgenticTyper surfaces 135 manually confirmed bugs and 40 possible bugs through increased type coverage that were previously unknown to the developers. A comparison with prior work shows that most existing tools are inapplicable or achieve at most 81.5% coverage while introducing thousands of type errors. Furthermore, a non-agentic LLM baseline fails to preserve runtime behavior.
The results validate that agentic AI enables automated typing at a practical level that prior approaches could not achieve, substantially reduces manual effort compared to fully manual typing, and identifies numerous existing issues. However, full automation is not achieved: human review remains essential for high-quality results, as manual cleanup reduces the number of added lines by 15-41% without sacrificing type coverage. Ultimately, the resulting pull requests, comprising 35,000 inserted lines, were merged by the development team, demonstrating the approach's practical value.
In everyday outpatient care, office-based medical professionals often face a growing volume of patient records, spanning years or even decades, while consultation times remain critically short. Much of this information is embedded in unstructured documents such as physician letters, lab reports, and diagnostic notes. The lack of time makes it difficult for physicians to comprehend the patients’ medical history and to base decisions on the full health and disease history. This paper presents a modular Visual Analytics approach particularly designed to support general practitioners in navigating long-term patient histories. Using transformer-based large language models, the system automatically extracts and classifies relevant clinical entities from unstructured documents. These extracted elements are mapped to a set of predefined data categories and visualized through tailored, interactive components, such as medication intake timelines, blood count diagrams, and named-entity highlighted texts. Visual components are selected dynamically based on content type and clinical relevance, allowing medical professionals to explore patient histories efficiently and to identify critical information at a glance. The modular architecture facilitates integrating new document types and visual modules, ensuring flexibility and scalability. Designed for real-world applicability, the system aims to reduce cognitive load, improve information accessibility, and support time-sensitive decision-making in primary care.
Featured Application
The ML Ontology presented in this article is the information backbone of a data science platform used for training machine learning models from datasets. Features include a training configuration wizard and interactive help.
This article presents an industry-ready ontology for the machine learning domain, which is named “ML Ontology”. While based on lightweight modelling languages, ML ontology provides novel features including built-in queries and quality assurance, as well as sophisticated reasoning. With ca. 700 individuals that define key ML concepts and ca. 5000 RDF triples, ML Ontology ranks among the largest domain-specific ontologies for ML. An experiment to estimate the correctness and completeness of ML terminology included in ML Ontology indicates an F1-score of 0.83. A benchmark evaluating query performance reveals query response times far below 100 ms even for complex queries and memory consumption below 3.5 MB. Its industry-readiness is demonstrated by benchmarks as well as two use case implementations within a data science platform. ML Ontology is open source and published under an MIT license.
Cardiovascular magnetic resonance imaging (CMR) offers state-of-the-art myocardial tissue differentiation. The CMR technique late gadolinium enhancement (LGE) currently provides the noninvasive gold standard for the detection of myocardial fibrosis. Typically, thresholding methods are used for fibrotic scar tissue quantification. A major challenge for standardized CMR assessment is large variations in the estimated scar for different methods. The aim was to improve quality assurance for LGE scar quantification, a multi-reader comparison tool “Lumos” was developed to support quality control for scar quantification methods. The thresholding methods and an exact rasterization approach were implemented, as well as a graphical user interface (GUI) with statistical and case-specific tabs. Twenty LGE cases were considered with half of them including artifacts and clinical results for eight scar quantification methods computed. Lumos was successfully implemented as a multi-level multi-reader comparison software, and differences between methods can be seen in the statistical results. Histograms visualize confounding effects of different methods. Connecting the statistical level with the case level allows for backtracking statistical differences to sources of differences in the threshold calculation. Being able to visualize the underlying groundwork for the different methods in the myocardial histogram gives the opportunity to identify causes for different thresholds. Lumos showed the differences in the clinical results between cases with artifacts and cases without artifacts. A video demonstration of Lumos is offered as supplementary material 1 . Lumos allows for a multi-reader comparison for LGE scar quantification that offers insights into the origin of reader differences.
Various types of anomalies can arise in cyber–physical production systems, caused by either faulty devices or incorrect processes. Anomalies within individual devices can often be detected by applying machine learning techniques to the respective produced multivariate time series. While this data typically shows temporal and spatial changes and can therefore be efficiently utilized by models, detecting anomalies within the process is often more challenging, as process data usually only consists of events, binary signals, or changes in unique process states. Due to the low variance of data, existing anomaly detection methods struggle to detect anomalies effectively and accurately. To address this challenge, in this paper, we propose a novel concept for process anomaly detection based on conditional discrete-time dynamic graphs. Through the conditional connections of the graph, essential characteristics can be generated and utilized to effectively train machine learning models to detect anomalies in the process data. Identified anomalies can be related to the current graph, facilitating transparent and explainable detections. By evaluating the concept against process data from an industrial unit and achieving an F1-Score of 0.96 and 1 for the realized repetitive processes, the accuracy and effectiveness of the concept can be demonstrated.
Background
Cardiovascular magnetic resonance (CMR) offers state-of-the-art volume, function, fibrosis and oedema imaging. Quality assurance (QA) tasks, such as quantitative parameter reproducibility assessments, the evaluation of AI methods, and the assessment of trainees have become essential to CMR. However, the explainability of how qualitative differences impact quantitative differences remains underexplored. Our aim is to demonstrate a semi-automated QA tool, Lazy Luna’s (LL) applicability to typical CMR QA application cases.
Methods
A software feature error-tracing is designed that allows for quickly pinpointing qualitative reasons for quantitative differences and outliers. Three QA application cases were designed. First, LL was applied to perform outlier detection for inter- and intraobserver analyses to detect failure cases and provide qualitative explanations. Outlier detection was performed on several typical images types. Second, LL supported an Artificial intelligence (AI) evaluation , in which an AI method was compared to a CMR-expert of 144 patients. LL assessed the acceptability of AI biases for left and right ventricular (LV, RV) end-systolic, –diastolic, and stroke volumes (ESV, EDV, SV), ejection fractions (EF) and the myocardial mass (LVM). Annotations were examined to explain the qualitative differences that resulted in good and poor parameters. The AI investigation was recorded as a video. Third, LL was used to provide a Trainee Feedback to a CMR beginner. The trainee was compared to an expert on several imaging techniques to investigate outliers.
Results
For the outlier detection, LL detected segmentation differences that caused parameter differences on multiple sequences. For the AI evaluation calculated clinical parameter biases to be: LVESV:-3.1 ml, LVEDV:2.1 ml, LVSV:6.5 ml, LVEF:3.0 ml, RVESV:0.3 ml, RVEDV:-3.8 ml, RVSV:-4.2 ml, RVEF:-1.4 ml, LVM:-2 g. Inspecting the causes for outlier differences revealed that juxtaposed basal slice failures caused unacceptable LVSV deviations between AI and expert. For the trainee assessment, LL showed that trainee parameters exceeded tolerance ranges. The segmentations could be improved to better mirror expert segmentations and close the parameter gaps.
Conclusion
Lazy Luna, as a semi-automated quality assurance tool, is applicable to several quality assurance application cases in CMR.
This article presents new trends and applications in generative artificial intelligence (AI) and knowledge-based AI. Foundations and applications of large language models (LLMs) are presented, such as classifying hate speech, easy-to-read language, and retrieval-augmented generation (RAG). The important aspect of user experience (UX) for AI systems is elaborated. Finally, AI solutions in different industry sectors are presented, in particular life sciences, the energy sector, and the manufacturing industry.
