Federated learning (FL)  is a technique to collaboratively train deep learning (DL) models without the need to host data on a central server. Instead, the training process occurs locally at each participating institution, and only model characteristics, e. g., weights or gradients, are shared to build a consensus model . Thus, as the data remains where it was acquired, FL resolves data governance and ownership issues, making it practical especially in medical imaging where data sharing may be critical due to the presence of personally identifiable information . Nevertheless, FL does not itself guarantee security and privacy since training data may be recovered by reconstruction attacks. Therefore, FL only offers an infrastructural approach to privacy and security, unless combined with other privacy-enhancing techniques (PETs) . Ziegler et al. investigated the feasibility of FL combined with a privacy mechanism called differential privacy (DP) [6, 7] for chest X-ray classification. However, the authors only conduct experiments with one aggregation method and a fixed number of clients.
In this work, we aim to explore FL approaches for the training of reliable thoracic abnormality classification systems. We will investigate different aggregation methods, e. g., [1, 8, 9], to be able to train the classifiers in a federated manner. In doing so, we will focus in particular on analyzing the effect of the number of clients. Furthermore, we will analyze the effect of additionally applied PETs such as DP on model training. We will perform an in-depth evaluation of the performance of the trained classifiers. For our experiments, we will employ the NIH ChestX-ray14  dataset, a collection of 112,120 frontal-view chest X-ray images from 30,805 unique patients with the text-mined fourteen abnormality labels.
The Master’s thesis covers the following aspects:
- Overview of the current state-of-the-art for PETs in medical imaging. This overview focuses in particular on FL and DP.
- Building a federated pipeline for training thoracic abnormality classification systems which includes:
- Hyper-parameter tuning
- Analyzing the effect of different aggregation methods
- Analyzing the effect of the number of clients
- Analyzing the effect of DP on model training
- Evaluating and comparing the performance of the trained classifiers.
- Investigating the feasibility and success of reconstruction attacks (optional).
All DL implementations will be implemented using PyTorch .
 Brendan McMahan et al. “Communication-Efficient Learning of Deep Networks from Decentralized Data”. In: Artificial Intelligence and Statistics. PMLR. 2017, pp. 1273–1282.
 Nicola Rieke et al. “The future of digital health with federated learning”. In: NPJ Digital Medicine 3.1 (2020), pp. 1–7.
 Kai Packhäuser et al. “Deep learning-based patient re-identification is able to exploit the biometric nature of medical chest X-ray data”. In: Scientific Reports 12.1 (2022), pp. 1–13.
 Georgios A Kaissis et al. “Secure, privacy-preserving and federated machine learning in medical imaging”. In: Nature Machine Intelligence 2.6 (2020), pp. 305–311.
 Joceline Ziegler et al. “Defending against Reconstruction Attacks through Differentially Private Federated Learning for Classification of Heterogeneous Chest X-Ray Data”. In: arXiv preprint arXiv:2205.03168 (2022).
 Cynthia Dwork et al. “Calibrating Noise to Sensitivity in Private Data Analysis”. In: Theory of Cryptography Conference. Springer. 2006, pp. 265–284.
 Alexander Ziller et al. “Medical imaging deep learning with differential privacy”. In: Scientific Reports 11.1 (2021), pp. 1–8.
 Hongyi Wang et al. “Federated Learning with Matched Averaging”. In: arXiv preprint arXiv:2002.06440 (2020).
 Tian Li et al. “Federated Optimization in Heterogeneous Networks”. In: Proceedings of Machine Learning and Systems 2 (2020), pp. 429–450.
 XiaosongWang et al. “ChestX-ray8: Hospital-Scale Chest X-Ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases”. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2017.
 Adam Paszke et al. “PyTorch: An Imperative Style, High-Performance Deep Learning Library”. In: Advances in Neural Information Processing Systems 32 (2019), pp. 8026–8037.