Machine Learning
Federated learning
Distributed learning scenarios where data remains close to its source while models or updates are coordinated across participants.
Research
Privacy-preserving machine learning over data that keeps changing.
My doctoral research combines privacy-preserving machine learning with data stream mining: learning from continuously arriving, non-IID data under concept drift, while keeping raw data protected through federated learning and homomorphic encryption.
Research themes
Privacy
Homomorphic encryption
Privacy-preserving computation patterns that enable selected operations over encrypted data, with explicit attention to performance tradeoffs.
Adaptive ML
Data stream mining
Learning from continuously arriving data under drift, latency, memory, non-IID distribution, and adaptation constraints.
Applied Research
Privacy-preserving ML in regulated domains
Research framing for environments where collaboration, confidentiality, and practical deployment constraints must coexist.
Current doctoral research
My doctoral thesis is titled Adaptive algorithms with privacy guarantees, supervised by Prof. Dr. Jean Paul Barddal and co-supervised by Prof. Dr. Vinícius M. A. de Souza at PUCPR's Graduate Program in Informatics, with funding from CAPES.
The work connects cryptographic and privacy-enhancing techniques with adaptive machine learning algorithms, for secure training and inference in distributed environments.
It considers supervised learning over continuous data streams: concept drift, non-IID data, incremental model adaptation, and the deployment constraints of regulated or data-sensitive domains.
Most privacy-preserving machine learning assumes a fixed dataset. Regulated enterprise data is online, drifting and resource-constrained, and that is the regime this research addresses.
Why it matters in practice
The organizations that would benefit most from machine learning are often the ones least able to centralize their data. Hospitals, banks, tax authorities and their suppliers hold information that is regulated, sensitive, and distributed across parties who cannot simply pool it.
Federated learning and homomorphic encryption address that directly: learning from data without collecting it, or computing over data without decrypting it. Both cost accuracy and processing time, and my work measures what they cost.
The part I find under-examined is time. Enterprise data arrives as a stream: its distribution shifts, and models degrade while still returning confident answers. Combining privacy guarantees with adaptation to drift is the harder problem, and the one closer to production.
Academic background and service
PhD candidate in Artificial Intelligence at PUCPR (2024—2028), following an MSc there in 2023 with a dissertation on peak prediction in time series applied to smart grids — the work behind the IJCNN 2024 paper — and a specialization in applied Artificial Intelligence. First degree in Computer Engineering from UEPG.
Teaches at PUCPR. Associate member of the Brazilian Computer Society (SBC) since 2024. Silver medal, Brazilian Informatics Olympiad.
Full professional proficiency in English and Spanish; native Portuguese. Citations appear as KRUGER, RODRIGO.
Research collaboration
I am interested in working with research groups on privacy-preserving machine learning over evolving data streams, including joint work and research exchanges during my doctoral program.