Federated Unlearning Without a Central Authority
Description: This PhD project investigates an open problem at the intersection of federated learning and machine unlearning: how can independent, decentralised parties agree to forget, correct, or discount a shared model contribution. Machine unlearning is well established for the single-owner case, and federated extensions are progressing fast. The project addresses two interconnected research challenges:
- Distributed Unlearning Theory
- Adversarial Robustness of the Unlearning Channel
Enrolment & opportunity: The successful candidate will enrol as a PhD student at the School of Computing Science under the supervision of Prof Christos Anagnostopoulos and will join the KDES Group. Our lab explores several different issues such as: distributed ML, statistical learning, scalable & adaptive information processing, and data processing algorithms.
Skills: The ideal candidate will have a background in computer science and a background in either mathematics and/or statistics. Special areas of interest include mathematical modelling/optimisation and distributed systems. A good understanding of basic machine learning and federated learning algorithms, as well as an MSc in one of the above areas, will be a considerable plus. Programming skills, a good command of English, and team-work capacity are required.
Contact: Prof Christos Anagnostopoulos
Distributed AI: Foundation Model Personalization
Research fields: Distributed ML; Foundational Models.
Description: This PhD project explores an emerging paradigm at the intersection of foundation models and federated learning: how can we efficiently personalize and fine-tune large pre-trained models across distributed mobile devices while respecting privacy constraints, computational limitations, and dynamic network conditions?
The proliferation of foundation models (large language models, vision transformers, multimodal models) has created unprecedented opportunities for personalized AI services. However, traditional fine-tuning approaches require either uploading sensitive user data to centralized servers or downloading massive models to resource-constrained devices. Neither approach is tenable for privacy-sensitive applications or bandwidth-limited mobile environments. This research investigates novel architectures and training protocols that enable collaborative model adaptation at the edge while preserving the capabilities learned during pre-training.
The project addresses four interconnected research challenges:
- Parameter-Efficient Federated Fine-Tuning
- Spatiotemporal Context-Aware Model Distribution
- Cross-Silo and Cross-Device Hybrid Architectures
Enrolment & opportunity: The successful candidate will enrol as a PhD student at the School of Computing Science under the supervision of Dr Christos Anagnostopoulos will join the KDES Group. Our lab explores several different issues such as: distributed ml, statistical learning, scalable & adaptive information processing, and data processing algorithms.
Skills: The ideal candidate will have a background in computer science and background in either mathematics and/or statistics. special areas of interest include: mathematical modelling/optimization. a good understanding of the basic machine learning and adaptation algorithms as well as an msc in one of the above areas will be a considerable plus. programming skills (python/matlab/java), good command of english and team work capacity are required.
Contact: Prof Christos Anagnostopoulos
Adaptive AI
Research fields: Adaptive AI & Distributed Large Models.
Description: This PhD project investigates fundamental challenges in deploying federated learning (FL) systems across mobile and edge computing infrastructures where clients exhibit unpredictable behavior, varying computational capabilities, and intermittent connectivity.
Traditional FL approaches assume stable client participation and reliable network conditions, which are assumptions that are rarely held in real-world mobile scenarios. This research will develop novel frameworks that intelligently adapt to client dynamics, optimize resource allocation under uncertainty, and ensure robust model training despite highly heterogeneous and volatile participation patterns.
The project encompasses three core research themes:
- Intelligent Client Selection Under Uncertainty Develop algorithms that predict and account for client availability, mobility patterns, and communication reliability when selecting participants for training rounds. This involves creating optimization frameworks that balance model quality objectives with practical constraints such as bandwidth limitations, completion probability, and fairness considerations.
- Decentralized Coordination Mechanisms Explore alternatives to centralized server architectures by investigating peer-to-peer aggregation protocols, hierarchical edge-cloud structures, and consensus-based coordination strategies. The research will examine how decentralized approaches can improve system resilience, reduce communication bottlenecks, and enable FL in scenarios where centralized coordination is impractical or undesirable.
- Privacy-Preserving Training with Heterogeneous Data Address the challenge of maintaining strong privacy guarantees while training over highly non-uniform data distributions across mobile devices. This includes investigating differential privacy mechanisms, secure multi-party computation protocols, and trusted execution environments that can operate efficiently on resource-constrained hardware without significantly degrading model performance.
The research will combine theoretical analysis (convergence guarantees, privacy bounds), algorithmic innovation (optimization under constraints), and empirical validation using real mobility traces and benchmark datasets. The ultimate goal is to bridge the gap between theoretical FL research and practical deployment requirements, creating systems that work reliably in production mobile environments rather than idealized experimental settings.
This project will contribute foundational knowledge on adaptive learning systems that can function effectively despite the inherent uncertainties and heterogeneities of mobile computing.
Enrolment & opportunity: The successful candidate will enrol as a PhD student at the School of Computing Science under the supervision of Dr Christos Anagnostopoulos will join the KDES Group. Our lab explores several different issues such as: distributed ml, statistical learning, scalable & adaptive information processing, and data processing algorithms.
Skills: The ideal candidate will have a background in computer science and background in either mathematics and/or statistics. special areas of interest include: mathematical modelling/optimization. a good understanding of the basic machine learning and adaptation algorithms as well as an msc in one of the above areas will be a considerable plus. programming skills (python/matlab/java), good command of english and team work capacity are required.
Contact: Prof Christos Anagnostopoulos