My research sits at the intersection of scalable learning systems, distributed intelligence, and trustworthy AI. I work on methods that are both scientifically grounded and useful in practice, spanning nested learning, federated learning and unlearning, agentic AI, efficient transformers, and physics-informed reinforcement learning.
TRL, IBA · Milestone
The Telecommunications Research Lab at IBA is growing
Our first Q1 publication is now out in IEEE Open Journal of the Communications Society, and around 20 more manuscripts are currently under review or in preparation across top-tier IEEE, ACM, and Q1 venues. It has been a genuine privilege to help build a research pipeline like this alongside Dr. Faisal Iradat and our international collaborators.
Hierarchical, multi-frequency learning frameworks that capture structure across timescales. Applied to federated optimisation, RL safety, medical imaging, ECG classification, environmental change detection, and 6G network slicing.
Federated Learning & Unlearning
Privacy-preserving distributed intelligence that handles client heterogeneity, catastrophic forgetting, and data-rights compliance, including selective knowledge removal without full model retraining.
Trustworthy & Safe AI
Safety architectures for human-AI interaction, including nested policy learning for developmentally adaptive child-AI safety and hard-constraint reinforcement learning with formal guarantees.
Agentic AI & LLMs
Multi-agent LLM pipelines for autonomous research and analysis, agentic financial analysis, skill-driven digital forensics orchestration, and cost-aware LLM routing (TokenGuard).
Efficient Transformers
Transformer architectures optimised for long-context modelling and irregular time-series data, with applications to cybersecurity, log analysis, and wearable-to-cloud continuum learning.
Applied ML for Security
Pre-encryption ransomware detection, explanation stability under temporal malware drift, LLM-based threat classification, and national-scale SIEM pipelines for government cyber-defence.
Publications
Click any paper to expand a description of its contribution.
JournalPublished
2026
Nested Multi-Agent Reinforcement Learning for Adaptive Resource Management in 6G Network Slicing: A Multi-Timescale Framework with Convergence Guarantees
A multi-timescale nested multi-agent RL framework for adaptive resource management in 6G network slicing, with convergence guarantees for the coupled slow/fast update processes. Targets the scalability-stability trade-off in dense 6G deployments. First Q1 publication from our TRL, IBA research programme.
Read more →
ConferenceAccepted
2026
Training-Free Coarse Point Cloud Registration via Axis-Projected Surface Area Matching
R. A. R. Ejaz, F. Iradat
ICCVDM 2026
Computer VisionPoint Clouds
A training-free coarse point cloud registration technique using axis-projected surface area matching, useful as a fast initial alignment stage before ICP-style refinement.
Read more →
ConferenceAccepted
2026
HADES: Hierarchical Autonomous Digital Evidence System Using Skill-Driven Multi-Agent Orchestration for Digital Forensics
R. A. R. Ejaz, Q. M. Waiz, F. Iradat
WI 2026, Linz (poster)
Agentic AIApplied ML for Security
HADES orchestrates skill-driven agents in a hierarchical forensic pipeline, automating evidence acquisition, triage, and correlation across heterogeneous artefacts while preserving chain-of-custody semantics.
Read more →
ConferenceAccepted
2026
From AI to Generative AI: How Emerging Technologies Are Reshaping Public Trust
E. Tariq, E. Tariq, I. S. Khan, R. A. R. Ejaz, F. Iradat
WI 2026 (poster)
Trustworthy AIAI & Society
Examines shifts in public trust as AI systems transition from discriminative to generative paradigms, drawing on a cross-region survey of perceived risk, utility, and institutional readiness.
Read more →
ConferenceAccepted
2026
Hierarchical Bayesian Nested Optimisation for Uncertainty-Aware Resource Allocation in 6G Open-RAN
R. A. R. Ejaz, Y. A. Bangash, F. Iradat, W. Iqbal
ICSL-DSGA 2026
Nested LearningBayesian Methods6G
Hierarchical Bayesian formulation of the Open-RAN resource-allocation problem with nested optimisation across control timescales, producing uncertainty-aware decisions under partial observability of network state.
Read more →
ConferenceAccepted
2026
TEMPEST: Temporal Execution Modeling for Pre-Encryption Ransomware Detection with Timestep-Level Explainability
A. Khan, A. Ahmedani, A. A. Kiani, R. A. R. Ejaz, Y. A. Bangash, et al.
ICSL-DSGA 2026
Applied ML for SecurityExplainability
TEMPEST models the temporal execution signature of ransomware to detect it pre-encryption, before any file is locked, while maintaining timestep-level explainability so analysts can see why each decision was made.
Read more →
JournalUnder Review
2026
It's the Method, Not the Model: Explanation Collapse Under Temporal Malware Drift
R. A. R. Ejaz, F. Iradat, W. Iqbal, M. Mansouri
IEEE Open Journal of the Communications Society
Trustworthy AIApplied ML for Security
Shows empirically that under temporal malware drift, explanation-method choice, not model choice, dominates explanation stability. Has implications for how we evaluate XAI in non-stationary security settings. Submitted to IEEE OJ-COMS.
Read more →
JournalUnder Review
2026
Nested Federated Learning: Layer-Wise Multi-Frequency Synchronization for Privacy-Preserving Distributed Intelligence
R. A. R. Ejaz, F. Iradat, W. Iqbal, M. Mansouri
Cognitive Computation
Federated LearningNested Learning
Introduces layer-wise multi-frequency synchronisation across federated clients, addressing scalability and client heterogeneity by decoupling which layers synchronise at which cadence.
Read more →
JournalUnder Review
2026
Multi-Frequency Layer-Wise Aggregation for Federated Clinical Prediction Across Hospitals
R. A. R. Ejaz, F. Iradat, M. I. Kharka, W. Iqbal
IJEHMC (IGI Global)
Federated LearningMedical Imaging
Layer-wise multi-frequency aggregation strategy for federated clinical prediction across hospitals, addressing distributional heterogeneity across institutions.
Read more →
JournalUnder Review
2026
Nested Learning for Adaptive-Fidelity ECG Classification: Matryoshka Representation Learning Across the Wearable-to-Cloud Continuum
R. A. R. Ejaz, F. Iradat, W. Iqbal, M. Kumail
IEEE Journal of Biomedical and Health Informatics
Nested LearningWearable ML
A Matryoshka-style nested representation scheme that allows ECG classifiers to run at reduced fidelity on wearables and expand smoothly to full-fidelity inference on cloud backends without separate model retraining.
Read more →
JournalUnder Review
2026
Nested Learning with Attention-Guided Multi-Frequency Supervision for Hepatic Vessel Segmentation
R. A. R. Ejaz, F. Iradat, W. Iqbal, A. Ahmad
Neural Computing and Applications (Q1)
Nested LearningMedical Imaging
Proposes a nested, attention-guided multi-frequency supervision scheme for segmenting hepatic vessels in medical CT. Hierarchical frequency decomposition with attention-gated fusion improves small-vessel recovery and boundary fidelity compared to single-scale baselines.
Read more →
ConferenceUnder Review
2026
DP-f-FUM: Differentially Private Federated Unlearning via Per-Sample Clipping and Min-Max f-Divergence Optimisation
R. A. R. Ejaz, F. Iradat, et al.
FLTA 2026 (main track)
Federated LearningTrustworthy AI
Differentially private federated unlearning combining per-sample clipping with min-max f-divergence optimisation for selective client removal with formal privacy guarantees.
Read more →
ConferenceUnder Review
2026
Active-Ledger: Trust-Gated Diffusion and Behavioural Contribution Scoring for Adversarially Robust Federated ECG Arrhythmia Classification
R. A. R. Ejaz, F. Iradat, et al.
AJCAI 2026
Federated LearningTrustworthy AI
Trust-gated diffusion combined with behavioural contribution scoring for adversarially robust federated ECG arrhythmia classification.
Read more →
ConferenceUnder Review
2026
FedSRIMU: Spectral Roughness-Informed Federated Unlearning for Privacy-Compliant IoT Intrusion Detection Systems
A. M. Warris, Z. U. Ebad, R. A. R. Ejaz, F. Iradat
UKCI 2026
Federated LearningApplied ML for Security
Spectral roughness-informed federated unlearning for privacy-compliant IoT intrusion detection systems. Companion undergraduate work I co-supervised.
Read more →
JournalUnder Review
2026
SafeNest: Nested Policy Learning for Multi-Timescale Safety in Child-AI Interaction
R. A. R. Ejaz, Y. A. Bangash, F. Iradat, M. Kumail
Scientific Reports
Trustworthy AIReinforcement Learning
SafeNest introduces nested policies where safety constraints are gated by developmental stage, enabling child-AI systems to adapt their interaction boundaries at multiple policy timescales rather than applying a single static filter.
Read more →
JournalUnder Review
2026
Nested Learning for Multi-Timescale Electric Vehicle Charging Coordination: A Physics-Informed Deep RL Framework with Hard Voltage Guarantees
R. A. R. Ejaz, N. Aburaed, F. Iradat
IEEE Transactions on Smart Grid
Reinforcement LearningNested LearningPhysics-Informed ML
A physics-informed deep-RL controller for large-scale EV charging that operates over nested timescales and provides hard voltage-constraint guarantees, bridging learned policies with provable grid-safety margins.
Read more →
ConferenceUnder Review
2026
NestAnt: A Nested-Learning Surrogate for Continual Multi-Band Microstrip Antenna Design
R. A. R. Ejaz, F. Iradat, W. Iqbal, Y. A. Bangash
IEEE GLOBECOM 2026 (ML for Communications & Networking)
Nested LearningApplied ML for Communications
Nested-learning surrogate for continual multi-band microstrip antenna design, targeting the transfer efficiency across neighbouring frequency bands.
Read more →
JournalIn Preparation
2026
Nested Federated Unlearning: A Multi-Frequency Optimization Framework for Privacy-Preserving Distributed Intelligence
R. A. R. Ejaz et al.
Target: IEEE Transactions on Artificial Intelligence
Federated LearningTrustworthy AI
A nested, multi-frequency optimisation framework enabling selective client-level knowledge removal in federated settings without full model retraining, a step toward compliance with data-rights legislation at scale.
Read more →
JournalIn Preparation
2026
A Nested Learning-Based Framework for Early Environmental Change Detection Using Multi-Temporal Remote Sensing Data
R. A. R. Ejaz et al.
Target: IEEE JSTARS
Nested LearningRemote Sensing
Applies nested multi-frequency learning to multi-temporal satellite imagery for early-warning detection of environmental change, targeting signals that appear well before they are visible to single-scale detectors.