Page 35 - Research and Innovation Pulse

TriGAN-SiaMT: A triple-segmentor adversarial network A systematic review of secure federated learning based on
with bounding box priors for semi-supervised brain lesion blockchain and Multi-Party computation

segmentation

Author(s):

Author(s):
Bhutta, M. N. M. (Abu Dhabi University), Irtaza, G. (University of Education, Lahore), Mehmood, A. (Abu Dhabi University),
Hamood, R. (National University of Science and Technology), Makhdoom, I. (National University of Science and Technology),
Alshurbaji, M. (Khalifa University of Science and Technology), Assefa, M. (Khalifa University of Science and Technology),
Obeid, A. (Khalifa University of Science and Technology), Seghier, M. L. (Khalifa University of Science and Technology), Elhadef, M. (Abu Dhabi University), Rehman, M. H. U. (King’s College London)
Hassan, T. (Abu Dhabi University), Taha, K. (Khalifa University of Science and Technology), Werghi, N. (Khalifa University of Index Terms:
Science and Technology)

Artificial intelligence; Collaborative learning; Distributed computer systems; Information leakage; Learning
Index Terms:
algorithms; Learning systems; Privacy-preserving techniques; Block-chain; Blockchain-based decentralization;
Decentralisation; Decentralised; Decentralized trust mechanism; Federated learning security; Integrity verifications;
Brain; Decision making; Image segmentation; Labeled data; Network architecture; Semi-supervised learning;
Teaching; Bounding-box; Brain lesion segmentation; Brain lesions; Deep learning; Exponential moving averages; Machine-learning; Model integrity verification; Privacy preserving; Privacy-preserving machine learning; Secure
Lesion segmentations; Mean-teacher; Siamese; Teachers’ multi-party computation; Trust mechanism; Blockchain

Abstract:
Abstract:

Federated Learning enables collaborative model training without compromising data privacy. However, security concerns
Accurate brain lesion segmentation in MRI is critical for clinical decision-making, but pixel-wise annotations remain costly
and time-consuming. We propose TriGAN-SiaMT, a novel semi-supervised segmentation framework that combines remain, particularly regarding participant contributions and model integrity. This paper explores the potential of integrating
adversarial learning, consistency regularization, and bounding box priors. Our architecture comprises three segmentors ( Blockchain and Multi-Party Computation techniques to address these challenges in Federated Learning. We systematically
S 0, S 1, S 2) and two discriminators ( D 0, D 1). It includes: (1) a supervised branch ( S 0↔ D 0) trained on a small labeled review recent research works on examining the capabilities of Blockchain-based Federated Learning and multi-party
subset; (2) a Siamese branch ( S 1↔ D 1) with an identical architecture to S 0↔ D 0, but trained on unlabeled data; and (3) a computation in mitigating security threats in federated learning, such as data leakage and model poisoning. In addition, by
teacher branch ( S 2) updated via exponential moving average (EMA) from S 1, following the Mean Teacher (MT) paradigm. analysing the convergence of these technologies, we aim to provide insights into their potential for building more secure,
The teacher S 2 generates pseudo-labels to supervise S 1. It also provides soft segmentations to guide D 1, which does trustworthy, and privacy-preserving Federated Learning. We conclude the review by identifying open research questions
not see any labeled data. The model enforces consistency at multiple levels: between S 0 and S 1 (Siamese consistency), and outlining promising directions for future research in this area, as this convergence is not only a technical achievement
and between S 1 and S 2 (EMA consistency). Bounding box priors are incorporated as weak supervision for both labeled but a foundational one towards democratised, secure, and privacy-aware artificial intelligence.
and unlabeled images, improving lesion localization. Evaluated on the ISLES 2022 and BraTS 2019 datasets, TriGAN-SiaMT Read the paper
achieves DSC scores of 84.80 % and 86.32 %, respectively, using only 5 % labeled data. These results demonstrate strong
performance under limited supervision and robust generalization across brain lesions.

Read the paper

Abu Dhabi University | Research and Innovation Pulse Newsletter 35