CDT Fort Cohort 1 complete Group Projects
The CDT Fort is proud to announce that the first cohort of students has successfully completed their group research projects.
Cohort 1 of the CDT Fort programme has successfully completed their group projects, marking a major milestone as they now transition into the research phase of their doctoral journey. The cohort of six researchers was divided into two cross-institutional teams, each exploring themes that emerged from the Ideas Lab hosted at Queen’s University Belfast in April, in collaboration with our industry partners. From May to August, the teams worked intensively to develop and deliver innovative solutions, culminating in the submission of a group conference paper alongside online presentations to an audience of academics, peers, and industry stakeholders. Project descriptions and abstracts from the resulting papers can be found below.
Project 1: ROBUST & RESILIENT SATELLITE COMMUNICATIONS
Satellite networks can support emergency responders in disaster situations where terrestrial communication networks are interrupted. In this situation, the resilience and security of the communications is imperative.
Strands:
WP1: Post quantum cryptography for satellite comms, Maxine Collins, Surrey
WP2: Distributed learning for satellite physical layer security, Saviz Changizi, Surrey
WP3: AI security for robust federated learning, Benjamin Gibney, QUB
Abstract:
This paper presents a modular framework for secure and resilient satellite communication in Low Earth Orbit (LEO), integrating post-quantum cryptography (PQC), anomaly detection, and trust-aware federated learning (FL). We implement a hybrid cryptographic stack combining Hamming Quasi-Cyclic (HQC) and dual-mode Advanced Encryption Standard (AES) for both confidentiality and side-channel resilience. To detect cyber–physical anomalies, we construct a synthetic telemetry dataset and evaluate lightweight models, finding that a combined convolutional neural network (CNN) and gated recurrent unit (GRU) architecture (CNN–GRU) achieves the best trade-off between accuracy and minority-class recall. We then deploy the anomaly detector in a federated learning setting and propose a trust-based aggregation method that gradually downweights poisoned updates. Compared to static defences, our approach maintains high accuracy even when malicious clients are in the majority. These results highlight the importance of a multi-faceted security strategy grounded in cryptography–machine learning (ML) integration for autonomous and trustworthy space systems.
Project 2: ORAN SMART NETWORKS
Towards an open network that leverages AI to improve efficiency while achieving a high level of security.
Strands:
WP1: AI/ML for network management, Chris Digby, Surrey
WP2: Trustworthy AI for network communications, Ousama Albagul, Surrey
WP3: Safe memory for AI applications, Yikang Shen, QUB
Abstract:
As attacks on the Internet of Things (IoT) increase, we urgently need IoT intrusion detection mechanisms. We fine tune the netFound foundation model on an 8-class subset of CICIoT2023 to evaluate IoT intrusion detection. Performance is significantly boosted by unfreezing a small number of hidden layers in the base model and by replacing the default classifier with a Random Forest. We then assess adversarial robustness under hardware and software attacks. At the hardware level, Rowhammer attacks can flip bits in serialized feature vectors: A bit-flip rate of 3.95% can cause a sample corruption rate of 71.8%. Our model F1 remains 0.79 under feature-vector poisoning, which is highly robust, whereas poisoning the training dataset causes F1 to collapse to 0.05. In parallel, we apply gradient-based attacks, Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD), on the trained model’s weights. Our results show that small perturbations can reduce accuracy by around 35%, while we further demonstrate that adversarial training improves robustness significantly.
Congratulations to the first cohort on this remarkable achievement, next stop PhD project!