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.
Research Areas
Nested Learning
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 benchmarking factual accuracy and hallucination rates.
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.
JournalUnder Review · 2026
Nested Multi-Agent Reinforcement Learning for Adaptive Resource Management in 6G Network Slicing: A Multi-Timescale Framework with Convergence Guarantees
R. A. R. Ejaz, F. Iradat, W. Iqbal
IEEE Open Journal of the Communications Society
Nested LearningReinforcement Learning6G
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.
Read more →
ConferenceUnder Review · 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 →
ConferenceUnder Review · 2026
TEMPEST: Temporal Execution Modeling for Pre-Encryption Ransomware Detection with Timestep-Level Explainability
R. A. R. Ejaz 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
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 →
JournalUnder Review · 2026
Nested Learning for Multi-Timescale Electric Vehicle Charging Coordination: A Physics-Informed Deep Reinforcement Learning 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 →
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 →
ConferenceUnder Review · 2026
A Multi-Timescale Safety Architecture for Developmentally Adaptive Child-AI Interaction via Nested Policy Learning
R. A. R. Ejaz, F. Iradat, M. Kumail
UKCI 2026
Trustworthy AINested Learning
Conference companion to SafeNest presenting the underlying multi-timescale safety architecture and empirical evaluation of developmentally adaptive constraint gating.
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, M. W. Iqbal, F. Iradat, M. Kumail
IEEE Xplore / IEEE Journal
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.
A layered detection framework that fires before encryption begins, combining behavioural, API-sequence, and file-entropy signals to stop ransomware earlier than post-encryption detectors.
Read more →
ConferenceUnder Review · 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
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 →
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
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.
Read more →
ConferenceUnder Review · 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
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 →
JournalUnder Review · 2026
The Impact of Female Under-representation in Cybersecurity on Women's Vulnerability to Cybercrime
M. S. Baig, R. A. R. Ejaz, F. Iradat
Gender, Technology and Development
AI & Society
A scoping analysis connecting workforce composition in cybersecurity to downstream vulnerability patterns, with policy recommendations for inclusive threat-modelling and support infrastructure.
Read more →
JournalIn Preparation · 2026
Nested Federated Unlearning: A Multi-Frequency Optimization Framework for Privacy-Preserving Distributed Intelligence
R. A. R. Ejaz et al.
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
Employment and Job Retention in the AI Era in LMICs, A Scoping Study
R. A. R. Ejaz, F. Iradat
TBD
AI & Society
Scoping study on how AI adoption is reshaping employment and job-retention patterns in low- and middle-income countries.
Read more →
JournalIn Preparation · 2026
Psychological and Ethical Factors Affecting Student Acceptance of AI-Generated Academic Work
R. A. R. Ejaz, F. Iradat
TBD
AI & Society
Empirical study of psychological and ethical factors driving student acceptance of AI-generated academic output, informing teaching-practice and policy design.
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.
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.
Read more →
JournalIn Preparation · 2026
Evaluating LLM-Generated Equity Research: A Benchmark Study of Factual Accuracy, Hierarchical Reasoning Quality, and Hallucination Rates in AI-Driven Financial Reporting
R. A. R. Ejaz et al.
TBD (MSc dissertation companion)
Agentic AILLMsBenchmarking
Open benchmark dataset and reproducible evaluation framework measuring factual accuracy, hierarchical reasoning quality, and hallucination rates of LLM-generated equity research against analyst-produced ground truth.