SKSaad Khan

Saad Khan

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Saad Khan has a total of 44 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    HuddsTrafficFL: A Benchmark Testbed and Dataset for Byzantine-Robust Federated Learning in Urban Traffic Systems

    by Farah Farid Babar, Saad Khan, Simon Parkinson

    Federated learning (FL) enables collaborative model training without sharing raw data, making it suitable for privacy-sensitive applications such as intelligent transportation systems and distributed Internet of Things (IoT) environments. However, evaluating Byzantine-resilient FL under realistic conditions remains challenging. Existing benchmarks mainly rely on image classification or generic network intrusion datasets and do not capture geographically distributed clients, non-IID data distributions, sensor-derived anomaly labels, and integrated Byzantine attack evaluation within a unified framework.To address this gap, we present HuddsTrafficFL, a reproducible benchmark dataset and evaluation framework developed from a city-scale traffic sensing environment using SUMO/TraCI simulations and OpenStreetMap road networks. HuddsTrafficFL provides three core contributions: (1) a geographically grounded FL dataset containing 91,482 timestep observations from 292 clients with natural non-IID characteristics; (2) sensor-derived anomaly labels generated from junction-level traffic behaviour, including sudden speed reductions and abnormal queue growth patterns; and (3) a configurable Byzantine attack and defence evaluation pipeline supporting reproducible robustness assessment across multiple datasets.To the best of our knowledge, HuddsTrafficFL is the first benchmark to integrate geographically grounded FL clients, sensor-derived anomaly labels, non-IID data distributions, and Byzantine attack evaluation within a single reproducible environment. We evaluate four Byzantine attacks and three robust aggregation methods across HuddsTrafficFL, CICIDS2017, and BoT-IoT. Results show that FedAvg degrades to an AUC of 0.50 under model poisoning, whereas coordinate median maintains an AUC of at least 0.9874 across all evaluated attack--dataset combinations. The benchmark resources will support reproducible research on secure federated learning for intelligent transportation and distributed sensing applications.

  • Open Access

    Wi-Vision: Domain-Adaptive Deep Learning for 3D Pose Estimationfrom WiFi Signals

    by Evans Aidoo, Minsi Chen, Saad Khan, Solomon Odoom, Zhijie Xu

    Most human pose estimation systems rely on cameras, raising privacy concerns and failing under occlusion or illumination changes. WiFi-based sensing is an attractive alternative, but existing methods lean on beamforming or phased arrays, restricting deployment to laboratory settings. We present Wi-Vision, a WiFi Channel State Information (CSI)-based framework operating on commodity single-antenna ESP32 hardware for privacy-preserving 3D human pose estimation. Wi-Vision employs a two-stage architecture: (1) a WiFi-to-2D heatmap encoder that maps CSI signals to joint probability distributions, achieving Percentage of Correct Keypoints at 5 cm (PCK@5cm) of 81.1% and PCK@10cm of 93.1% across 7 diverse subjects, and (2) a 2D-to-3D lifting network trained on WiFi-predicted poses fused with monocular depth estimates. Unlike prior camera-dependent methods, our domain-adaptive training strategy generates 3D ground truth directly from WiFi-predicted 2D poses, ensuring train-test distribution consistency and achieving an 8.3-fold improvement overcamera-trained baselines. Extensive experiments on 7 participants comprising 376,689 synchronized samples yield a Mean Per-Joint Position Error (MPJPE) of 146.1 mm (104.0 mm excluding head joints), operating entirely on commodity hardware during inference without any camera or specialized radio-frequency (RF) equipment. Our domain-adaptive approach establishes strong benchmarks for RF-based pose estimation on commodity hardware, opening pathways for ubiquitous sensing in smart homes, healthcare, and security.

  • Open Access

    Reproducibility in event-log research: a parametrised generator and benchmark for event-based signatures

    by Saad Khan, Simon Parkinson, Monika Roopak

    Event-based datasets are crucial for cybersecurity analysis. A key use case is detecting event-based signatures, which represent attacks spanning multiple events and can only be understood once the relevant events are identified and linked. Analysing event datasets is essential for monitoring system security, but their growing volume and frequency create significant scalability and processing difficulties. Researchers rely on these datasets to develop and test techniques for automatically identifying signatures. However, because real datasets are security-sensitive and rarely shared, it becomes difficult to perform meaningful comparative evaluation between different approaches. This work addresses this evaluation limitation by offering a systematic method for generating event logs with known ground truth, enabling reproducible and comparable research. We present a novel parametrised generation technique capable of producing synthetic event datasets that contain event-based signatures for discovery. To demonstrate the capabilities of the technique, we provide a benchmark in signature detection. Our benchmarking demonstrated the suitability of DBSCAN, achieving a score greater than 0.95 Adjusted Rand Index on most generated datasets. This work enhances the ability of researchers to develop and benchmark new cybersecurity techniques, ultimately contributing to more robust and effective cybersecurity measures.

  • Open Access

    Fine-Tuning Small Language Models for Solution-Oriented Windows Event Log Analysis

    by Siraaj Akhtar, Saad Khan, Simon Parkinson

    Large language models (LLMs) have shown promise for event log analysis, but their high computational requirements, reliance on cloud infrastructure, and security concerns limit practical deployment. In addition, most existing approaches focus only on the identification of the problem and do not provide actionable remediation. Small language models (SLMs) present a lightweight alternative that can be fine-tuned for a specific purpose and hosted locally. This paper investigates whether SLMs, when fine-tuned for a specific task, can serve as a practical alternative for event log analysis while also generating solutions. We first create a large-scale synthetic Windows event log dataset that contains remediation actions using a high-performing LLM. We then fine-tune multiple SLMs and LLMs using the LoRA parameter-efficient fine-tuning technique and evaluate their performance by comparing with expert assessment. The results show that the dataset accurately reflects real-world scenarios and that fine-tuned SLMs consistently outperform LLMs in identifying issues and providing relevant remediation, while requiring fewer computational resources.

  • Open Access

    Asymmetric Angular Margin Learning for Open-Set Face Presentation Attack Detection

    by Mohammadreza Sheikhfathollahi, Simon Parkinson, Saad Khan

    Face presentation attack detection (PAD) is essential for securing face recognition systems against diverse spoofing attacks, such as printed photographs, replayed videos, and 2D/3D mask-based manipulations. Most deep learning–based PAD approaches formulate the task as a closed-set binary classification problem, which restricts their ability to generalise to unseen attack types encountered in real-world scenarios. To address this limitation, we propose an asymmetric angular margin–based discriminative learning framework tailored for open-set PAD. Unlike conventional angular margin losses (e.g., ArcFace) that impose symmetric compactness across all classes, our formulation introduces an additive angular margin exclusively for the live class within a softmax-based optimisation scheme. This asymmetric design enforces tight hyperspherical clustering of live embeddings while allowing spoof representations to remain distributed, thereby better reflecting the intrinsic heterogeneity of spoof attacks. By integrating angular normalisation, learnable class centres, and scaled cross-entropy optimisation, the proposed approach reshapes the decision boundary from a linear hyperplane to an angular hyperspherical separation mechanism. Extensive experiments on benchmark datasets, including Oulu-NPU, CASIA-MFSD, Replay-Attack, MSU-MFSD, and SiW-Mv2, demonstrate that the proposed method achieves superior generalisation to unseen attacks, consistently outperforming conventional binary classifiers and symmetric angular margin baselines in terms of HTER, and AUC. These results validate the effectiveness of asymmetric angular discriminative learning for robust open-set face anti-spoofing.

  • Open Access

    Reproducibility in Event-Log Research: A Parametrised Generator and Benchmark for Event-based Signatures

    by Saad Khan, Simon Parkinson, Monika Roopak

    Event-based datasets are crucial for cybersecurity analysis. A key use case is detecting event-based signatures, which represent attacks spanning multiple events and can only be understood once the relevant events are identified and linked. Analysing event datasets is essential for monitoring system security, but their growing volume and frequency create significant scalability and processing difficulties. Researchers rely on these datasets to develop and test techniques for automatically identifying signatures. However, because real datasets are security-sensitive and rarely shared, it becomes difficult to perform meaningful comparative evaluation between different approaches. This work addresses this evaluation limitation by offering a systematic method for generating event logs with known ground truth, enabling reproducible and comparable research. We present a novel parametrised generation technique capable of producing synthetic event datasets that contain event-based signatures for discovery. To demonstrate the capabilities of the technique, we provide a benchmark in signature detection. Our benchmarking demonstrated the suitability of DBSCAN, achieving a score greater than 0.95 Adjusted Rand Index on most generated datasets. This work enhances the ability of researchers to develop and benchmark new cybersecurity techniques, ultimately contributing to more robust and effective cybersecurity measures.