Robust, explainable and privacy-aware machine learning systems for distributed and adversarial security settings.
Trustworthy AI is not a single metric. In security-sensitive systems it combines robustness, explainability, privacy, accountability and evidence that a model behaves acceptably under operational stress.
Distributed AI systems inherit all the usual machine learning risks and add communication constraints, partial observability, heterogeneous peers and inconsistent incentives. Trustworthy design must address the model, the protocol and the operational context together.
The related publications and projects approach trustworthy AI through robust aggregation, decentralized protocols, privacy-preserving learning, explainable mitigation support and evaluation in cyberdefense scenarios.
A trustworthy AI system is robust, explainable, privacy-aware, auditable and evaluated under conditions close to the environment where it will be used.
No. Explainability is important, but it must be combined with robustness, privacy, evaluation, governance and operational controls.
Federated systems rely on multiple clients and shared updates. Trustworthy design helps detect unreliable contributions, protect sensitive information and keep collaborative learning useful.
DEFENDIS develops a framework for uniquely identifying IoT devices in a distributed manner while solving security threats through decentralized federated learning.
View ProjectROBUST-6G studies security mechanisms for 6G systems, including monitoring, secure data management, trustworthy AI services, federated learning, and threat response.
View ProjectA project with armasuisse Cyber-Defence Campus investigating trustworthy threat intelligence sharing and verification algorithms.
View ProjectDeveloped decentralized AI solutions addressing computational constraints and trust validation under armasuisse Cyber-Defence Campus sponsorship.
View ProjectMove from foundations to design and evaluation. Each note addresses a different question.
Learn how federated learning and FedAvg work, what differential privacy and secure aggregation protect, and how to evaluate collaboration.
Combine differential privacy, secure aggregation and robust DFL with explicit trust assumptions, local threat budgets and joint evaluation.
Evaluate DFL robustness with F1, false rejection, overhead and recovery. Interpret Flighter and compare clean and attacked conditions.
Open questions in robust DFL: dynamic graphs, manipulated context, privacy, missing modalities and experiments to test the limits.
This site loads optional analytics from Google and external analytics providers only if you accept. You can decline and continue using the site normally.