Representation and Resource-Aware Federated Learning for Real-World Systems

Abstract

My research focuses on building trustworthy, decentralized intelligence   AI systems that can learn from data distributed across many devices, organisations, or sensors without ever centralising that data. I work mainly in federated and distributed learning, designing methods that cope with real-world messiness: non-IID data, missing or weak labels, asynchronous clients, stragglers, and strict privacy and resource constraints. A core theme in my work is enhancing data representation in the federated setting  using representation learning, clustering, and semi-supervised techniques so that global models remain robust and generalizable even when clients see very different data and only some of it is labelled.

Beyond the core algorithms, I am also interested in how these ideas play out on edge and hardware-constrained systems, such as smartphones, IoT devices. I explore resource-aware and post-deployment adaptation (e.g., pruning, quantisation, and personalised fine-tuning under device budgets) and apply federated learning to domains like smart cities. Overall, my goal is to develop FL frameworks that are not only accurate, but also privacy-preserving, resource-efficient, and deployable in complex real-world environments.

Contact: Tahani Aladwani

Publications

 

 

GraphChain: Temporal Subgraph Learning for Cyber Attack Path Reconstruction

Abstract

Modern software and AI supply chains are increasingly targeted by sophisticated and multi-stage attacks, yet existing detection systems achieve limited effectiveness due to insufficient feature engineering and a lack of models capable of capturing such complex attack patterns. These threats propagate across heterogeneous components and data sources, making it essential to integrate multi-modal, distributed telemetry to obtain a coherent view of system state. To address this challenge, temporal graph–based methods offer a principled way to model evolving dependencies, encode subgraph-level behaviors, and reveal subtle anomalies that unfold over time. Leveraging temporal graph representations thus provides a more accurate and scalable foundation for detecting advanced supply-chain attacks in dynamic environments. 

Contact: Zhuoran Tan

Publications

  • Z. Tan, S. P. Parambath, C. Anagnostopoulos, J. Singer, and A. K. Marnerides, “Advanced persistent threats based on supply chain vulnerabilities: Challenges, solutions, and future directions,” IEEE Internet of Things Journal, vol. 12, no. 6, pp. 6371–6395, 2025.
  • Z. Tan, C. Anagnostopoulos, and J. Singer, “Osptrack: A labeled dataset targeting simulated execution of open source software,” in 2025 IEEE/ACM 22nd International Conference on Mining Software Repositories (MSR), 2025, pp. 659–663.
  • Z. Tan, C. Anagnostopoulos, S. P. Parambath, and J. Singer, “Unified semantic log parsing and causal graph construction for attack attribution,” 2024. [Online]. Available: https://arxiv.org/abs/2411.15354.
  • Tan, Z., Anagnostopoulos, C., & Singer, J. (2025). Distributed temporal graph learning with provenance for APT detection in supply chains. arXiv. https://arxiv.org/abs/2504.02313

Knowledge distillation in heterogeneous federated learning

Abstract

In heterogeneous federated learning problems, knowledge distillation strategies are introduced to address client drift issues arising from local data or model heterogeneity. Furthermore, knowledge distillation aligns the performance of client models.

Key words: Distributed machine learning, knowledge distillation, federated learning.

Contact:   Wenhao Li

 

Publications 

 

Data Driven Approaches for Detecting False Data Injection Attacks toward Energy Market

Abstract

My research focuses on understanding and detecting stealthy manipulation of electricity markets, particularly attacks that subtly alter Locational Marginal Prices (LMPs). I work on building realistic datasets that model how attackers can manipulate grid parameters, loads, or topology without triggering traditional bad-data detectors. These datasets enable reproducible experimentation and support the development of more reliable detection methods. In parallel, I design lightweight, adaptive detection models capable of operating under real market uncertainty. This includes a drift-aware unsupervised anomaly detector and an incremental LSTM-based method that identifies abnormal price behaviour over time. My current work also involves modelling the energy market as a graph signal processing problem for clustering and node selection, as well as developing system-wide drift detection techniques.

Contact: Ghadeer Alsharif

Publications: 

  •  G. O. Alsharif, C. Anagnostopoulos and A. K. Marnerides, "Energy Market Manipulation via False-Data Injection Attacks: A Review," in IEEE Access, vol. 13, pp. 42559-42573, 2025, doi: 10.1109/ACCESS.2025.3548914.
  • G. O. Alsharif, C. Anagnostopoulos and A. K. Marnerides, "Incremental Learning Detection of Distributed Financially Motivated Attacks in Energy Markets," in Proc. 2025 IEEE 45th Int. Conf. on Distributed Computing Systems Workshops (ICDCSW), pp. 189–194, 2025, doi: 10.1109/ICDCSW63273.2025.00038.
  • G. O. Alsharif, C. Anagnostopoulos and A. K. Marnerides, "Drift-aware Unsupervised Detection of Stealthy FDIA Towards Energy Market," in Proc. 2025 IEEE Int. Conf. on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm), pp. 1–7, 2025, doi: 10.1109/SmartGridComm65349.2025.11204629. 
  • G. O. Alsharif, C. Anagnostopoulos, A. K. Marnerides and P. Mathaios, "SMLT: A Synthetic Dataset for Stealthy Manipulation of Energy Market via False Data Injection Attacks," (Accepted in ECMLPKDD Workshop 2025.)

 

Safe & Scalable Multi-Agent Coordination via Level-k Game Theory & Tree Search

Abstract

This project builds a family of modular, structure-aware intelligent engines designed for multi-agent systems operating in uncertain, safety-critical environments. The system integrates recursive game-theoretic reasoning, multi-objective cost modeling, and safety-constrained tree search to enable realistic interactions between human-driven and autonomous vehicles. While initially developed for autonomous driving, these engines, such as our KAN-LSTM model and multi-objective MCTS-Level-k planner, offer generalizable structures for decentralized decision-making, uncertainty reasoning, and trajectory optimization. This mechanism enables real-time, trustworthy decision-making in dynamic environments, while offering a theoretical foundation for robust planning and coordination in collaborative logistics and sensor-integrated intelligent systems.

Contact: Zhihao Lin

Publications: 

 Project Website: 

 

Federated Learning and Optimization in Mobile Edge Computing Environments

Abstract

Mobile Edge Computing (MEC) environments present unique challenges for distributed machine learning due to client mobility, heterogeneous resources, intermittent connectivity, and varying network conditions. Our research focuses on developing intelligent frameworks for federated learning (FL) that operate effectively in dynamic mobile environments, addressing fundamental challenges in client selection, resource allocation, and system optimization.
 
A key contribution in this area is MOBILE (Mobility and Outage-Based Intelligent Federated Learning), which challenges conventional assumptions that prioritize energy constraints over mobility patterns. By leveraging historical mobility data and formulating client selection as a regularized Mixed-Integer Quadratic Programming (MIQP) problem, MOBILE achieves up to 89% successful client participation rates compared to 32% with standard approaches, while reducing wasted throughput from 73% to 31%. Our work addresses broader challenges in MEC-based FL systems including optimizing client selection under mobility uncertainty, efficient bandwidth allocation, handling non-i.i.d. data distributions, and minimizing resource waste while preserving user incentives. This research is conducted as part of the EU Horizon Europe ELLIE project.

 

Contact:  Qiyuan Wang

Publications

  • Qiyuan Wang, Qianyu Long, and Christos Anagnostopoulos. "MOBILE: Mobility and Outage-Based Intelligent Federated Learning in Mobile Computing." In ECAI 2025 - 27th European Conference on Artificial Intelligence, I. Lynce et al. (Eds.), Frontiers in Artificial Intelligence and Applications, Vol. 413, pp. 2977-2984. IOS Press, 2025. doi: 10.3233/FAIA251158

 

 

Resilient Inference for Personalized Federated Learning in Edge Environments

Abstract

Personalized FL (PFL) enhances traditional FL by tailoring models to the needs of each participant and improving local inference performance. Existing methods often ignore the impact of the inherent model differences on inferential tasks leading to suboptimal personalization in EC. To address this challenge, we propose a framework, SOIR, which effectively integrates model similarity into the rescheduling process. The inference task rescheduling problem in EC is formulated with a Mixed Integer Nonlinear Programming model, and an efficient algorithm is introduced to solve it. Inspired by this work, our future research will build upon the relevant findings in this paper to explore topics including FL security, thereby enhancing the resilience of FL.

Contact:  Ke Xiao

Publucations