Projects

Modern vision–language models (VLMs) such as CLIP, BLIP-2, and LLaVA  have shown remarkable capabilities in understanding images and text jointly. However, these multimodal AI systems face critical security vulnerabilities along two intertwined dimensions: robustness and privacy. In high-stakes applications like cyber-defence, such vulnerabilities could be disastrous. An adversary might subtly perturb an image to mislead (...)

National digital identity systems, namely Swiss e-ID and EUDI, are going live in 2027. For such a system to be usable, it must provide account recovery, accountability, and uniqueness, without sacrificing user privacy. Cryptographic solutions to each have existed for decades, but they fail in deployment. The work in this fellowship hence co-designs cryptographic protocols (...)

Deployed cyber and cyber-physical systems need to evolve to respond to newly discovered threats and maintain confidentiality, integrity, and availability. In many cases, this evolution takes the form of patches: small changes to an already-deployed system. Timely patching is crucial, but dangerous: bad patches can weaken or break a system. In high-criticality systems, full system (...)

Secure electronic vaults protect sensitive data, but users risk permanent loss when credentials are forgotten. Unlike Apple social recovery, this approach enables self-recovery while preserving trustee privacy. This allows to serve customers seeking trustworthy, privacy-first storage solutions.

Evaluate the needs for Digiinov to comply to CRA concerning their existing Human Machine Interface (HMI) product. Research how to interface this HMI with an LLM, what security implications it has, and how local and cloud-based LLMs fare wrt performance, security, and CRA.

This project aims to systematically study security risks in foundation models, with a specific focus on backdoor threats.

This project aims to develop scalable and principled methods for forecasting, interpolation, and anomaly detection in dynamic networks and multivariate time series by combining optimal transport (OT), graph structure learning, and spatiotemporal neural models.

Building a modular and extensible framework designed to harden consumer drone firmware against attacks. The framework targets five key areas: firmware rehosting for analysis, peripheral modeling to simulate drone hardware, communication protocol evaluation and hardening, verification of trusted firmware components and execution environments, and analysis of Android-based controller applications. Together, these components establish a comprehensive (...)

The project will evaluate and validate visual AI methods for interpreting diagrams, figures, and charts in scientific documents. It aims to strengthen technology-monitoring and early-warning capabilities by detecting emerging innovations, research trends, and weak signals that text-based analysis may miss. The work includes developing a capability map of existing methods and piloting a selected method (...)

The project will demonstrate how search-augmented language models can support technology monitoring through temporally controlled access to external knowledge. By restricting searches to selected time periods, analysts can track how technologies evolve, detect emerging trends and inflection points, and monitor the spread of technical concepts without retraining the model. The approach provides traceable sources and (...)

The project will assess the feasibility of using Multi-Agent Systems to support armasuisse’s technology-monitoring activities. It will examine how access to different internal and external data sources affects output quality, runtime, and cost. The project will also produce two to three sample technology-monitoring reports through an iterative process incorporating feedback and expert insights from armasuisse.

Today's identity security faces challenges like misuse and tracking. Our goal is to enable secure, anonymous, unlinkable E-ID interactions by researching novel cryptographic algorithms. This boosts user trust, creates new business opportunities, and cuts financial losses after data breaches.

The project developed and matured an automated security-testing solution for IoT software that combines firmware rehosting with feedback-guided fuzzing. By decoupling embedded software from its physical hardware, the solution creates standardized, scalable virtual testing environments that can automatically explore software behavior, identify security-relevant anomalies, and provide concrete inputs that help engineers reproduce and fix vulnerabilities. (...)

Focus on AI models that can be self-hosted (open weights) and are trustworthy as by having open training data, and have enhanced multilingual capabilities in languages relevant for the Swiss community.

This project proposes to design metrics, methods, and scalable algorithms for detecting anomalies in dynamic networks. The temporal evolution of the structure of dynamic networks carries critical information about the development of complex systems in various applications, from biology to social networks. Deviations from regular network structure evolution may also provide critical information about anomalies (...)

One of the increasingly popular paradigms for managing the growing size and complexity of modern ML models is the adoption of collaborative and decentralized approaches. While this has enabled new possibilities in privacy-preserving and scalable frameworks for distributed data analytics and model training over large-scale real-world models, current approaches often assume a uniform trust-levels among (...)

Modern infrastructure management increasingly relies on infrastructure-as-code (IaC), which is the paradigm of automatically managing computing infrastructure using programming languages such as Ansible. Furthermore, there is an increasing interest to leverage Large Language Models (LLMs) to 1) automatically generate the specification code that provisions the desired infrastructure, and 2) to periodically check if the infrastructure (...)

State-of-the-art architectures in many software applications and critical infrastructures are based on deep learning models. These models have been shown to be quite vulnerable to very small and carefully crafted perturbations, which pose fundamental questions in terms of safety, security, or performance guarantees at large. Several defense mechanisms have been developed in the last years (...)

Phishing is a growing societal threat and requires urgent, effective solutions. Through a collaborative process involving field feedback, ideation workshops, and state-of-the-art analysis, this open innovation project will explore technical, organizational, and legal solutions. The results will be prototypes, practical guidelines, and recommendations to strengthen digital security against phishing.

DISCO-DHRIVE is developing a privacy-preserving collaborative learning platform using AI. It allows the building of AI models across different locations without the need to share sensitive data. Tailored to meet ICRC's unique challenges, including resource scarcity and stringent data confidentiality, the project integrates federated and distributed learning. This approach enables the extraction of valuable insights from sensitive data without compromising their security.

Pre-trained foundation models are widely used in deep learning applications due to their advanced capabilities and extensive training on large datasets. However, these models may have safety risks because they are trained on potentially unsafe internet-sourced data. Additionally, fine-tuned specialized models built on these foundation models often lack proper behavior verification, making them vulnerable to adversarial attacks and privacy breaches. The project aim is to study and explore these attacks in for foundation models.

The main objective of the project is to perform online monitoring of technologies and technology actors in publicly accessible information sources. The monitoring concerns the early detection of mentions of new technologies, of new actors in the technology space, and the facts related to new relations between technologies and technology actors (subsequently, all these will be called technology mentions). The project will build on earlier results obtained on the retrieval of technology-technology actors using Large Language Models (LLMs).

The temporal evolution of the structure of dynamic networks carries critical information about the development of complex systems in various applications, from biology to social networks. While this topic is of importance, the literature in network science, graph theory, or network machine learning, still lacks of relevant models for dynamic networks, proper metrics for comparing network structures, as well as scalable algorithms for anomaly detection. This project exactly aims at bridging these gaps.

The slowdown in Moore’s Law has pushed high-end GPUs towards narrow number formats to improve logic density. This introduces new challenges for accurate Deep Neural Network (DNN) training and inference. Our research aims to bring novel solutions to the challenges introduced by ubiquitous ever-growing DNN models and datasets. Our proposal targets building DNN platforms that are optimal in performance/Watt across a broad class of workloads and improve utility by unifying the infrastructure for both training models and inference tasks.