Publications

Publications:  Karol Kozak‬ - ‪Google Scholar‬ 

Wir löschen die Daten. Das Modell erinnert sich trotzdem. Die stille Ironie der Künstlichen Intelligenz in Verwaltung und Praxis 

ENG: We Delete the Data. The Model Remembers Anyway. The Quiet Irony of Artificial Intelligence in Administration and Practice

"Wir löschen die Daten. Das Modell erinnert sich trotzdem." dieser Satz klingt zunächst wie ein technisches Detail, fast wie eine Randnotiz aus der Welt der Informatik. Tatsächlich beschreibt er sehr präzise das Spannungsfeld, in dem sich öffentliche Verwaltungen beim Einsatz von Künstlicher Intelligenz (KI) heute bewegen. Zwischen Effizienzgewinn, wachsendem Druck zur Digitalisierung und der gleichzeitigen Verantwortung für sensible Daten entsteht eine neue Realität, die sich nicht mehr mit den klassischen Kategorien der Informationstechnologie (IT) erklären lässt. 

eng: “We Delete the Data. The Model Remembers Anyway.” At first glance, this sentence sounds like a technical detail, almost a side note from the world of computer science. In reality, however, it describes with remarkable precision the tension in which public administrations find themselves today when deploying artificial intelligence (AI). Between gains in efficiency, growing pressure to digitize, and the simultaneous responsibility to protect sensitive data, a new reality is emerging that can no longer be fully explained through the traditional categories of information technology (IT).

Robust Machine-vision models against adversarial attacks in Industry 4.0

Background Industrial production systems increasingly depend on deep learning–based machine-vision models for automated inspection, quality assurance, and process monitoring in Industry 4.0 environments. While these models achieve high accuracy under normal conditions, they are vulnerable to adversarial attacks that introduce subtle, visually imperceptible perturbations. Such attacks can lead to defect masking, misclassification, and potentially unsafe system behaviour. Despite extensive research on adversarial robustness in generic vision tasks, systematic studies tailored to industrial machine-vision pipelines remain limited. Methods This work analyses adversarial threats against industrial machine-vision systems and proposes a novel defence mechanism, Consistency-based Machine-Vision Adversarial Defence (C-MVAD). The approach is designed as a drop-in extension for convolutional neural networks and combines adversarial training, lightweight regularisation, and embedding-consistency checks. Robustness is evaluated against three gradient-based attacks, namely Fast Gradient Sign Method (FGSM), Basic Iterative Method (BIM), and Projected Gradient Descent (PGD). Experiments are conducted on an industrial defect detection benchmark using a compact convolutional backbone suitable for edge deployment. Performance is assessed in terms of classification accuracy under attack, adversarial detection capability, and runtime overhead. Results The baseline model exhibits a substantial accuracy degradation under BIM and PGD attacks, with accuracy dropping below 30 % for PGD at moderate 

Secure3D-CV: A scientific framework for integrity assessment of 3D data with an external Python library

Three-dimensional (3D) perception is widely used in robotics, autonomous vehicles and medicine. These systems often operate autonomously and are vulnerable to data tampering or spoofing attacks. Existing work on 2D adversarial detection does not generalise well to the high-dimensional and structured nature of 3D point clouds.

We propose Secure3D-CV, a method that builds on OpenCV by adding a library for outlier detection and threat assessment in 3D scenes. The method transforms each scene ( S) into a feature vector ( x = φ(S)) via geometric, spectral and volumetric descriptors. A statistical anomaly score ( ) measures the deviation of ( x) from a benign distribution ({ }). A supervised or one-class classifier ( f θ ( x)) estimates ( P( y = 1| x)) where ( y) denotes a safe scene. The final decision score ( R( S)) strategically combines ( A( S)) and ( f θ ( x)) to classify the scene as safe or not safe.

On two datasets, simulated point clouds and depth maps containing 2,000 samples, and 100 anonymised CT volumes, we show that Secure3D-CV detects injected outliers, occlusions and adversarial perturbations with high true-positive rates (99.3% on synthetic data and 98.7% on CT data) and low false-positive rates

Lightweight NSM: A clinical-linguistic approach to narcissism detection on social media

The rise of social media has brought renewed attention to the study of personality traits like narcissism. While much research has focused on platforms like Facebook and Instagram, decentralized networks such as Mastodon remain relatively unexplored. This paper introduces Lightweight NSM, a text-based model for detecting narcissistic traits in social media posts, grounded in the clinical criteria of the DSM-5. We collected and annotated a dataset of 2,000 Mastodon posts using GPT-4.1-mini, finding that a threshold of two or more DSM-5 criteria provided a reasonable baseline for identifying potentially narcissistic content. Our model, a simple Logistic Regression classifier using TF-IDF features, was trained on this annotated dataset. The model achieved an F1-score of 0.69 for the narcissistic class. Our analysis revealed a notable difference in narcissistic content between posts tagged with #selfie (29.1%)

A Dependable Explainable NLP Framework for Clinical Personality Trait and Subtype Analysis

The application of natural language processing models in clinical and mental health contexts requires not only high predictive performance but also interpretability, robustness, and well-calibrated uncertainty. This paper presents a dependable and explainable NLP framework for fine-grained personality trait analysis, grounded in clinical theory. The proposed approach combines a parameter-efficient DistilBERT architecture with lightweight adapter modules, resulting in only 1.91% trainable parameters, thereby reducing overfitting risks in domain-specific datasets. To address severe class imbalance, the framework integrates data augmentation with a Focal Loss objective. Explainability is achieved through a multi-layer approach combining local post-hoc explanations with class-level statistical validation. The framework is evaluated on a hold-out test set of 1,600 social media posts, achieving 82.06% accuracy 

PS-SecureScan: Pre-Execution Security Analysis for Photoshop Image Workflows

Image files are commonly treated as passive digital assets; however, modern formats such as PSD, TIFF, JPEG and PNG are complex structured containers that require extensive parsing, decompression and metadata interpretation. In creative production environments Adobe Photoshop is one of the most widely used tools for processing such files, making it a potential entry point for image-based cyberattacks. Vulnerabilities in image parsing routines, metadata handling and decompression mechanisms have historically enabled remote code execution, denial-of-service attacks and covert data exfiltration.

This paper analyzes cybersecurity risks associated with image processing in Photoshop-centered workflows and introduces PS-SecureScan, a pre-execution image examination framework designed to detect malicious or anomalous image files before they are opened in production software. The framework integrates structural validation, metadata anomaly detection, entropy-based steganography analysis, decompression risk estimation and sandboxed behavioral observation. Experimental evaluation using synthetic benign and adversarial image samples demonstrates that the proposed multi-layer analysis approach enables effective differentiation between normal creative assets and manipulated images. The results suggest that pre-execution image examination can significantly reduce the risk of image-based compromise in digital media workflows.

Book Cybersecurity for Love

The most precious thing in life is often our well-being, encompassing physical health, mental balance, and true happiness. And at the core of happiness lies one of life’s most powerful forces—love. Love brings connection, belonging, and emotional fulfillment that can light up our existence, whether it’s found in romantic relationships, deep friendships, or the bonds of family. Love, in all its beautiful forms—romantic, familial, or platonic—nourishes the soul. It’s what makes us feel alive, valued, and complete. But what happens when the digital world we live in threatens that love? As we venture deeper into the age of cyberattacks, even love is not immune from the risks. Cyberattacks don’t just target systems—they target lives, and when they hit close to home, they can strain even the strongest bonds. Imagine the emotional toll of a privacy breach, where your personal moments are exposed, or the financial strain a cyberattack could cause. Trust issues can emerge, stress levels rise, and relationships may falter under the weight of these new challenges. Yet, the flipside is just as powerful: overcoming these obstacles together can strengthen relationships, fostering resilience and deepening your connection with those you love. So, dear reader, embark with me on a thrilling journey through the world of cybersecurity and love—where digital threats meet human emotion. In a world where we’re more connected than ever, protecting love means not only shielding your heart but also your digital life. Because perfect love, even in the face of digitalization’s fears, has the power to cast out that fear. Welcome to a celebration of the individual, their love, and the profound impact cybersecurity can have in safeguarding what we hold most dear.