Research

Trustworthy, distributed, efficient.

SAIL organizes its work around three pillars connecting AI security, privacy, federated systems, and practical deployment.

The page is intentionally editorial and lightweight: no search widgets, no dense sidebars, and no dark background blocks.

Research pillars

Three areas frame our current work and connect to projects and publications.

Trustworthy AI

We study how machine learning systems behave under uncertainty, distribution shifts, adversarial conditions, and privacy constraints. Our work focuses on robustness, privacy protection, backdoor attacks and defenses, trustworthy evaluation, and safer deployment.

RobustnessPrivacyBackdoor attacks and defensesModel reliabilityExplainabilityFederated securityTrustworthy evaluation
Explore pillar

Distributed Learning

We design learning systems that work across distributed clients, data silos, institutions, and edge devices without centralizing private data. Our work includes federated learning, personalized learning, federated unlearning, fairness, communication efficiency, and cross-silo collaboration.

Federated learningCross-silo learningPersonalized federated learningFederated unlearningClient heterogeneityFairnessPrivacy-preserving collaboration
Explore pillar

Efficient Machine Learning

We build efficient AI systems that reduce computation, communication, memory, and deployment cost. Our work studies resource-constrained learning, edge AI, efficient training, lightweight architectures, low-rank methods, and green AI infrastructure.

Communication efficiencyEdge AIResource-constrained learningLow-rank trainingEfficient inferenceGreen AILightweight architectures
Explore pillar

Trustworthy AI

We study how machine learning systems behave under uncertainty, distribution shifts, adversarial conditions, and privacy constraints. Our work focuses on robustness, privacy protection, backdoor attacks and defenses, trustworthy evaluation, and safer deployment.

Representative questions

  • How can AI systems remain reliable under data shifts and adversarial conditions?
  • How should privacy and utility be balanced in sensitive ML applications?
  • How can backdoor risks be measured and mitigated in distributed settings?
RobustnessPrivacyBackdoor attacks and defensesModel reliabilityExplainabilityFederated securityTrustworthy evaluation

Selected publications

Clean-Label Physical Backdoor Attacks with Data Distillation · 2026BackFed: A Standardized and Efficient Benchmark Framework for Evaluating Backdoor Attacks in Federated Learning · 2026Wicked Oddities: Selectively Poisoning for Effective Clean-Label Backdoor Attacks · 2025Backdoor Attacks and Defenses in Federated Learning: Survey, Challenges and Future Research Directions · 2024
active

TrustFed: Trustworthy Federated Large Language Models

A research project on trustworthy federated learning for large language models, focusing on robustness, privacy, evaluation, and scalable collaboration.

Funded by the Accelerating Research Excellence Program, VinUniversity. Principal Investigator: Prof. Kok-Seng Wong.

Timeline: 2026–2028

Federated learningLarge language modelsTrustworthy AIPrivacyRobustness
Related papers
  • FedDDF: Dynamic Dataset Filtering in Federated Large Language Model Training
active

Privacy-Preserving, Robust, and Explainable Federated Learning for Healthcare

Federated learning methods for healthcare systems where privacy, robustness, and interpretability are central requirements.

Healthcare AIPrivacyRobustnessExplainabilityCross-silo learning
Related papers
  • Personalized Privacy-Preserving Framework for Cross-Silo Federated Learning
  • On the Trade-off Between Privacy Protection and Data Utility for Chest X-ray Images

Distributed Learning

We design learning systems that work across distributed clients, data silos, institutions, and edge devices without centralizing private data. Our work includes federated learning, personalized learning, federated unlearning, fairness, communication efficiency, and cross-silo collaboration.

Representative questions

  • How can multiple institutions learn together without centralizing private data?
  • How can federated systems adapt to heterogeneous devices, clients, and data distributions?
  • How can distributed systems onboard new clients while retaining previous knowledge?
Federated learningCross-silo learningPersonalized federated learningFederated unlearningClient heterogeneityFairnessPrivacy-preserving collaboration

Selected publications

HFedATM: Hierarchical Federated Domain Generalization via Optimal Transport and Regularized Mean Aggregation · 2026Onboarding Without Forgetting: Hypernetwork Personalization with Data-Free Replay for Personalized Federated Learning · 2026BackFed: A Standardized and Efficient Benchmark Framework for Evaluating Backdoor Attacks in Federated Learning · 2026An Empirical Study of Federated Learning on IoT-Edge Devices: Resource Allocation and Heterogeneity · 2025
active

TrustFed: Trustworthy Federated Large Language Models

A research project on trustworthy federated learning for large language models, focusing on robustness, privacy, evaluation, and scalable collaboration.

Funded by the Accelerating Research Excellence Program, VinUniversity. Principal Investigator: Prof. Kok-Seng Wong.

Timeline: 2026–2028

Federated learningLarge language modelsTrustworthy AIPrivacyRobustness
Related papers
  • FedDDF: Dynamic Dataset Filtering in Federated Large Language Model Training
active

Privacy-Preserving, Robust, and Explainable Federated Learning for Healthcare

Federated learning methods for healthcare systems where privacy, robustness, and interpretability are central requirements.

Healthcare AIPrivacyRobustnessExplainabilityCross-silo learning
Related papers
  • Personalized Privacy-Preserving Framework for Cross-Silo Federated Learning
  • On the Trade-off Between Privacy Protection and Data Utility for Chest X-ray Images

Efficient Machine Learning

We build efficient AI systems that reduce computation, communication, memory, and deployment cost. Our work studies resource-constrained learning, edge AI, efficient training, lightweight architectures, low-rank methods, and green AI infrastructure.

Representative questions

  • How can learning systems reduce communication and computation while remaining accurate?
  • How should AI models be adapted for edge devices and constrained environments?
  • How can training infrastructure become more scalable and resource-efficient?
Communication efficiencyEdge AIResource-constrained learningLow-rank trainingEfficient inferenceGreen AILightweight architectures

Selected publications

HFedATM: Hierarchical Federated Domain Generalization via Optimal Transport and Regularized Mean Aggregation · 2026Memory-efficient Continual Learning with Prototypical Exemplar Condensation · 2026An Empirical Study of Federated Learning on IoT-Edge Devices: Resource Allocation and Heterogeneity · 2025SC-GIR: Goal-oriented Semantic Communication via Invariant Representation Learning for Image Transmission · 2025
active

TrustFed: Trustworthy Federated Large Language Models

A research project on trustworthy federated learning for large language models, focusing on robustness, privacy, evaluation, and scalable collaboration.

Funded by the Accelerating Research Excellence Program, VinUniversity. Principal Investigator: Prof. Kok-Seng Wong.

Timeline: 2026–2028

Federated learningLarge language modelsTrustworthy AIPrivacyRobustness
Related papers
  • FedDDF: Dynamic Dataset Filtering in Federated Large Language Model Training
active

Green Serverless Computing for Resource-Efficient AI Training

Resource-efficient AI training infrastructure with an emphasis on greener, scalable serverless computing.

Green AIResource efficiencyServerless computingEfficient training
Related papers
  • Memory-efficient Continual Learning with Prototypical Exemplar Condensation