AMLD 2021 Workshop – Federated Learning: collaborative machine learning on sensitive decentralized data
Event details
| Date | 29.09.2021 |
| Hour | 09:00 › 13:00 |
| Location | |
| Category | Conferences - Seminars |
⚠️ A valid COVID certificate must be presented on site to enter the event. ⚠️
Datasets of interest in many application domains (e.g. healthcare, finance data, manufacturing) contain sensitive or private information and cannot easily be shared. Additionally, such data frequently belongs to multiple distinct parties and combining it in one location would expose a lucrative target to hackers. Therefore, it is desirable to make use of such data without a need to disclose it or store it in a central location.
Unfortunately, traditional methods to train predictive models expect data to be fully accessible and centralized on a single server. Research work therefore has to rely on small or artificial datasets that can safely be centralized. As a result, findings frequently do not generalize well to real-world datapoints and progress is hampered.
Federated Learning (FL) is a recently introduced paradigm that addresses this limitation by training models on decentralized datasets without requiring centralized data access. This approach allows multiple distinct parties to collaboratively train predictive models without a need to directly share sensitive data. Instead of combining datasets, FL trains a model in multiple iterations on data subsets stored in different locations. In every iteration, every party owning a data subset downloads a copy of the current model weights. An updated model is computed for each data subset in a local training step. Per-party model updates are then aggregated (a step that can be centralized, as it does not require data access) resulting in a single overarching FL update step.
As an introduction to the workshop, we will introduce the basic concepts underlying FL and discuss a few of the key related topics (e.g. Differencial Privacy,
Model Encryption). Our focus however, will be on gaining hands-on experience. We will implement a simple Federated Learning system using tensorflow (tensorflow/federated) and pytorch (PySyft). We will give a quick introduction to all needed libraries and tools at the start.
Datasets of interest in many application domains (e.g. healthcare, finance data, manufacturing) contain sensitive or private information and cannot easily be shared. Additionally, such data frequently belongs to multiple distinct parties and combining it in one location would expose a lucrative target to hackers. Therefore, it is desirable to make use of such data without a need to disclose it or store it in a central location.
Unfortunately, traditional methods to train predictive models expect data to be fully accessible and centralized on a single server. Research work therefore has to rely on small or artificial datasets that can safely be centralized. As a result, findings frequently do not generalize well to real-world datapoints and progress is hampered.
Federated Learning (FL) is a recently introduced paradigm that addresses this limitation by training models on decentralized datasets without requiring centralized data access. This approach allows multiple distinct parties to collaboratively train predictive models without a need to directly share sensitive data. Instead of combining datasets, FL trains a model in multiple iterations on data subsets stored in different locations. In every iteration, every party owning a data subset downloads a copy of the current model weights. An updated model is computed for each data subset in a local training step. Per-party model updates are then aggregated (a step that can be centralized, as it does not require data access) resulting in a single overarching FL update step.
As an introduction to the workshop, we will introduce the basic concepts underlying FL and discuss a few of the key related topics (e.g. Differencial Privacy,
Model Encryption). Our focus however, will be on gaining hands-on experience. We will implement a simple Federated Learning system using tensorflow (tensorflow/federated) and pytorch (PySyft). We will give a quick introduction to all needed libraries and tools at the start.
Links
Practical information
- General public
- Registration required