Home  |  Selected Publications  |  Talks  |  Service  |  Teaching  |  Project  |  Honors  |  Google Scholar  |  GitHub

Publications

Using programmable network devices to aid in-network machine learning has been the focus of significant research. This work presents Planter, an open-source, modular framework for mapping trained machine learning models to programmable devices. Planter supports a wide range of machine learning models, multiple targets and can be easily extended.

Deliverable 4.1 is the first iteration of the Dynamic Swarm Networking technical Work Package. The Work Package combines three tasks: (1) Automatic discovery and dynamic network swarm formation (2) Embedded network security and isolation and (3) Hardware accelerated in-network operations.

This work presents IIsy, which implements machine learning classification models in a hybrid fashion using off-the-shelf network devices. Besides a range of traditional and ensemble machine learning models, IIsy also supports hybrid classification, achieving near-optimal classification results, while significantly reducing latency and load on the servers.

In-network caching expedites data retrieval by storing frequently accessed data items within programmable data planes, thereby reducing data access latency. In this paper, we introduce In-network Caching Shelter (INCS), an in-network machine learning solution implemented on NVIDIA BlueField-2 DPU to mitigate the effect of bots’ traffic. Our evaluation shows that INCS can detect malicious bot traffic patterns with high accuracy.

In-network ML is a promising technology that provides ML inference services with high throughput, low latency and high power efficiency. This paper provides a holistic review of the fundamentals of in-network ML. It introduces its background and solutions, discusses the existing challenges and open issues, and provides insights for further research explorations.

Sparse-reward reinforcement learning environments pose a particular challenge because the agent receives infrequent rewards, making it difficult to learn an optimal policy. In this paper, we propose NSSE, a novel approach that combines stratified state space exploration with prioritised sweeping to enhance the informativeness of learning, thus enabling fast learning convergence.

Research has focused on enabling on-device functionality, with limited consideration to distributed in-network computing. This paper explores the applicability of distributed computing to in-network computing and presents DINC, a framework enabling distributed in-network computing, generating deployment strategies, overcoming resource constraints and providing functionality guarantees across a network.

Current gateways struggle with dynamic IoT traffic and have limited defence capabilities against attacks with changing patterns. In this work, we present P4Pir, a novel in-network traffic analysis framework for IoT gateways. P4Pir pioneer the continuous and seamless updates of in-network inference models within gateways.

Traffic load balancing is a long-time networking challenge. This work presents QCMP, a ReinforcementLearning based load balancing solution implemented within the data plane, providing dynamic policy adjustment with quick response to changes in traffic. Our results show that QCMP requires negligible resources, runs at line rate, and adapts quickly to changes in traffic patterns.

IoT devices are widely and dynamically deployed in diverse use cases, leading to a surge of security threats that were previously overlooked. We propose P4Pir, an in-network traffic analysis solution, to enable continuous learning and consistent ML updates. P4Pir achieves real-time multi-protocol data collection, in-network ML-based attack mitigation, and hitless runtime ML updates.

The rise of IoT-connected devices has led to an increase in collected data for service and traffic analysis, but also to emerging threats and attacks. To address these concerns, we present FLIP4, a distributed in-network attack detection framework based on federated tree models.

Machine learning is driving the evolution of algorithmic trading, but the demands for fast execution speed remain. In this paper, we present LOBIN, providing machine learning based market prediction by building limit order books within programmable switches, using high-frequency market data feeds.

In this paper, we introduce a novel history-aware epsilon-greedy logic for hyperparameter optimisation that adjusts the hyperparameters at runtime based on the analysis of the agent’s performance over time windows in a single agent’s lifetime. Our experimental results show that the proposed framework significantly improved performance compared to traditional hyperparameter tuning approaches.

We propose an adversarial autoencoder method to reduce MRI variations in AD, which is robust to variations that exist in multi-center MRI-scans in AD classification. The proposed method gives better performance than baseline approaches and provides a fully automatic end-to-end approach for AD classification.

Many trading applications require very short response times, which cannot always be supported by traditional machine learning frameworks. We present Linnet, providing in-network financial market prediction. Linnet demonstrates the potential to predict future stock price movements with high accuracy and low latency.

Traditional hard-coded gateways fail to flexibly process diverse IoT traffic from highly dynamic devices. We present P4Pir, an in-network traffic analysis solution for IoT gateways. Preliminary results show that P4Pir can detect emerging attacks accurately based on retraining and updating the machine learning model.

Appropriate vulnerability management is difficult to achieve when applying deep RL in a real-life problem. We propose a framework that allows users to focus on the algorithm of reasoning, trust, and explainability in accordance with human perception, followed by exploring potential threats, especially adversarial attacks and countermeasures.

Using programmable network devices to aid in-network machine learning has been the focus of significant research. This work presents Planter, an open-source, modular framework for mapping trained machine learning models to programmable devices. Planter supports a wide range of machine learning models, multiple targets and can be easily extended.

This work presents IIsy, which implements machine learning classification models in a hybrid fashion using off-the-shelf network devices. Besides a range of traditional and ensemble machine learning models, IIsy also supports hybrid classification, achieving near-optimal classification results, while significantly reducing latency and load on the servers.

This work presents ETeMoX, an architecture based on temporal models to keep track of the decision-making processes of RL systems. The architecture also integrates complex event processing, an event-driven approach, for detecting matches to event patterns that need to be stored, instead of keeping the entire history.

Data classification within the network significantly benefits reaction time, servers offload and power efficiency. This work designs an algorithm for efficient mapping of ensemble models, such as XGBoost and Random Forest, to programmable switches. The proposed method overlaps trees within match-action tables, achieves high accuracy and low resource overhead.

This work proposes the MNN-CH, a novel modular neural network that is constructed with explored category hierarchy. The basic idea is learning to learn an optimized category hierarchy to decompose complex patterns. And specific patterns are imposed into corresponding modules to realize a transparent design of the neural network.

Achieving an optimal global goal remains a persistent obstacle for collaborative multiagent systems. This work develops a reward-reinforced generative adversarial network to represent the distribution of the value function, replacing the approximation of Bellman updates. We demonstrated our method is resilient and outperforms other conventional reinforcement learning methods.

Unconstructed data mining is a challenging task in the current data mining research domain. In this work, we examine a large number of tables that existed in the bioenergy field and propose a framework for automatically acquiring data from papers in PDF format and outputting the table information.

Tongue carcinoma is one of the most common oral and maxillofacial malignant tumors worldwide, maximum standardized uptake value has been widely used in cancer research. This study aims at investigating the relationship between them and whether SUV parameters can predict lymph node metastasis.

This study investigates the relationship between maximum standardized uptake value (SUVmax) of ¹⁸F-FDG PET/CT and clinicopathological features of oral squamous cell carcinoma (OSCC), in order to formulate a better clinical guideline.

Finding a way within a given maze is a common problem for mobile robots. Existing routing algorithms usually can not able to find the solution to a filled path. This work presents a new path routing approach for mobile robots to get an intuitive filled path, which can solve the difficulty of modelling and be well applied to realistic problems.

Functional and structural brain connectivity in the triple network is consistently altered in schizophrenia and major depressive disorder. This study proposes the SCNMF to extract distributed multi-modal brain patterns to distinguish schizophrenia and major depressive disorder in a latent low-dimensional space of the triple brain network.

In allusion to the phenomenon that the multipath value of BeiDou MEO satellite of its regional system varies with the elevation angle of the observation station, we design and implemented an in-orbit test on a BeiDou test satellite. The result shows that the influence of the multipath effect of BeiDou global system is less than 0.1 m.