Professor Claudio Orlandi

In our digital era, centralized systems pose a significant risk, susceptible to failures and targeted by malicious actors. A breach in these systems can result in devastating consequences for data confidentiality, integrity, and availability. The growing interest in decentralized cryptographic solutions, like Secure Multiparty Computation, Blockchains, and Threshold Cryptography, arises from a resilience against both external and internal threats. However, deploying these systems reveals a mismatch between theoretical assumptions and real-world scenarios, potentially leading to severe security consequences.

The DECRYPSYS project addresses this gap, aiming to design cryptographic systems with robust theoretical foundations. Focused on aligning abstract models with the dynamic reality of deployment, especially in decentralized systems and the challenges posed by quantum computing, the outcome aims to produce novel, efficient, and secure cryptographic protocols. All project results will be publicly accessible, contributing to the safeguarding of critical digital infrastructures for businesses and society.

Professor Kasper Green Larsen

Machine learning's transformative impact now extends across science, industry, and society. Within the TUCLA project, we're advancing our understanding of classic machine learning, focusing on key algorithms like Bagging and Boosting. Deep learning thrives at tasks like generative AI, image- and speech recognition, in particular when data and compute resources are abundant. Classic algorithms shine when data is more structured, tabular, or scarce. Our goal is to enhance classic machine learning, offering faster algorithms demanding less training data for accurate predictions. Three key objectives drive our project:

Innovative Learning Framework: Establishing a novel theoretical framework, we'll prove generalization bounds for learning algorithms. New Boosting algorithms will accurately predict with less training data, complemented by proof that no other algorithm can do better.

Parallel Boosting Algorithms: Designing parallel Boosting versions enables training with computationally expensive base learning algorithms, potentially also elevating Boosting's role in deep learning.

Bagging's Theoretical Applications: Recent breakthroughs reveal Bagging's optimal status in classic learning setups. We'll explore theoretical applications of Bagging in other crucial settings, contributing to our understanding and development of classic machine learning.