Der Vortrag beschäftigt sich damit, wie sich neuronale Netze sicher nutzen und gleichzeitig ihre Entscheidungen erklären lassen. Dafür wird die sichere Inferenz mittels Mehrparteienberechnung (MPC) mit der Erklärbarkeitsmethode Occlusion Analysis kombiniert. Es wird gezeigt, wie sich diese Methode in eine sichere Inferenz integrieren lässt und wie sich die Qualität der Erklärungen sowie der zusätzliche Rechenaufwand im Vergleich zu einer herkömmlichen Inferenz verhalten.

Die Teilnehmenden erhalten einen Einblick in die Herausforderungen sicherer und erklärbarer KI. Sie erfahren, dass sich mit der Occlusion Analysis auch im MPC-Setting aussagekräftige Erklärungen erzeugen lassen, dies jedoch mit einem deutlich höheren Rechenaufwand verbunden ist. Dabei wird deutlich, dass Datenschutz und Erklärbarkeit miteinander vereinbar sind, ihre praktische Umsetzung jedoch zusätzliche Ressourcen erfordert.

Oblivious pseudorandom functions (OPRFs) are a cryptographic primitive that lets two parties jointly evaluate a pseudorandom function without revealing their secret inputs. OPRFs enable privacy-preserving services such as checking passwords against leaked databases and secure password-based login. Companies such as Cloudflare, Google, and Apple deploy protocols with OPRFs. However, most practical OPRFs can be broken by future quantum computers.

Isogeny-based cryptography is a promising post-quantum approach with small key sizes and low communication overhead. Several isogeny-based OPRFs have been proposed, broken, and repaired or replaced by new constructions, but a systematic comparison is still lacking.

The presented thesis addresses this gap by providing the first comprehensive overview and unified analysis of isogeny-based OPRFs. The existing constructions are presented, categorized by their underlying building blocks, and compared by computational, round, and communication complexity. In addition, a unified security analysis is provided, enabling direct comparison of security guarantees across protocols. The result is a structured reference and foundation for future research.

Smartwatches und Fitness-Tracker erfassen Vitaldaten, Bewegungen und Aufenthaltsorte rund um die Uhr und können in Ermittlungsverfahren entscheidende Beweise liefern. In der deutschen Ermittlungspraxis werden sie dennoch weitgehend vernachlässigt. Der Vortrag stellt die Ergebnisse von sechs Experteninterviews mit Vertreterinnen und Vertretern aus Behörden, Wissenschaft und Privatwirtschaft vor und zeigt: Die zentrale Hürde ist nicht primär technischer Natur, sondern ein Awareness-Defizit.

Die Teilnehmenden lernen drei aus der Praxis gewonnene Kernprinzipien kennen: das Drei-Quellen-Prinzip, das Wearable, Companion-App und Cloud gemeinsam betrachtet und die App als erfolgversprechendsten Zugangsweg priorisiert, die Sieben-Tage-Regel für das kritische Zeitfenster der Beweissicherung und die in der Praxis entwickelte WAMPIRE-Methode, die ein Wearable-Problem in ein Smartphone-Forensik-Problem überführt. Ein Vier-Phasen-Modell und eine Priorisierungs-Matrix zeigen, wie sich Wearable-Forensik ohne eigene Spezialeinheit in bestehende Abläufe der Mobile-Forensik integrieren lässt.

Modern mobile operating systems restrict access to explicit hardware identifiers. Tracking entities increasingly exploit benign operational data for device fingerprinting. To quantify this attack surface, DroidSonar, a dynamic API-scanning tool, is deployed across 24 physical smartphones running Android 11 through 16. DroidSonar calls all APIs it can detect during its runtime. Using a filtering pipeline, the analysis identified 513 highly stable, zero-permission APIs that can be exploited for fingerprinting, averaging 230.8 APIs per device. The empirical evaluation reveals that the Android permission model does not offer sufficient protection against fingerprinting. Furthermore, original equipment manufacturer customizations exacerbate the issue; for example, devices from Samsung, OPPO, OnePlus, and Xiaomi expose numerous manufacturer-exclusive tracking endpoints. OS Feature bloat causes the number of usable fingerprintable APIs to increase steadily from Android 11 to 16. Chaining minimal subsets of hardware properties with user-configured states reliably isolates individual devices, underscoring the need for operating-system-level attribute normalization to prevent fingerprinting.

Cellular networks provide connectivity on the go, enabling remote access to various services, including emergency services, but the deployed protocol stacks are highly complex. In smartphones, baseband chipsets implement them and provide the device's primary application processor with connectivity. However, most of such chipsets lack common security features, making them easy targets for over-the-air exploitation. In 2025, Apple debuted the iPhone 16e with the first iteration of its in-house baseband, the C1. Until now, there has been only limited research into basebands in Apple's platforms and no review of their security mitigations.

We survey the hardware schematics of the C1 baseband, examine its iOS integration, and reverse-engineer its security model for memory and control-flow integrity. The C1's security mitigations are comparable to those of modern application processors. We also assess its processing of untrusted over-the-air inputs by comparing 112 firmware data structures against 3GPP standards.

Lockdown Mode is an optional security mode developed to protect at-risk users against targeted attacks. To provide enhanced protection, it restricts or disables selected system functionality. These restrictions make Lockdown Mode an interesting subject of analysis, offering insights into which attack vectors and system components Apple considers critical to security. Comparing operating system versions and platforms, such as macOS and iOS, also reveals how this assessment evolves over time. However, while Apple documents the restrictions at a high level, their implementation, cross-platform behavior, and evolution remain largely opaque. Academic research on Lockdown Mode is also limited.

This talk presents the results of a technical analysis of Apple’s Lockdown Mode on seven macOS and one iOS version using static and dynamic reverse engineering methods. It introduces three mechanisms for querying Lockdown Mode status and shows which system components are affected. Furthermore, it describes selected examples from the 75 identified features whose behavior is modified by Lockdown Mode, comparing their implementation across macOS and iOS versions.

Modern Face Recognition (FR) models require large amounts of facial image data for training. However, data protection concerns and strict legal guidelines severely limit the collection of real datasets. Synthetic face generation using Diffusion Models has established itself as a privacy-friendly alternative, yet existing approaches still face the challenge of balancing identity consistency, identity separability, and intra-class variety.

This talk presents two novel diffusion-based approaches. The first, ABP, relies on attribute-based prompts; a prompt describes a phenotypic class rather than one exact identity. The second, M-SDFR, replaces text conditioning with identity embeddings from FRoundation, a biometric foundation model used in place of the text encoder. FR models trained on M-SDFR data achieve up to 96.35% on LFW and 86.21% average accuracy, outperforming GAN-based methods and remaining competitive with State-of-the-Art models trained on smaller-scale datasets. Although a gap to methods trained on larger datasets remains, foundation-model-based identity conditioning provides a robust basis for generating FR training data that complies with data protection regulations.

The spread of AI-generated fake news and deepfakes has increased the need for reliable ways to verify the trustworthiness of digital content. Content Credentials, developed by the Coalition for Content Provenance and Authenticity (C2PA), address this challenge by recording a medium’s provenance and using cryptographic hash functions and digital signatures to protect its authenticity. But do these mechanisms deliver the security guarantees they promise?

In this talk, I formalize authenticity for C2PA Content Credentials through a game-based definition and provide a corresponding security proof. I demonstrate that, given suitable cryptographic schemes, forgery is limited to negligible-probability events involving hash collisions or signature forgery. However, I also present a potential weakness in the hard binding, which links a Content Credential to its associated medium using a cryptographic hash. For practical reasons, the C2PA specification allows certain data ranges to be excluded from hash generation. I show that this exclusion reduces the hash input space, potentially enabling collisions and undermining the authenticity of Content Credentials.