1 of 85

AIDA Technical Day

Federated Learning

1

2025-09-16, Erik Ylipää (erik.ylipaa@liu.se)

2 of 85

AIDA Technical Days

September 16 - Federated Learning

13:00 - 13:50 - Introduction to Federated Learning

13:50 - 14:00 - Break

14:00 - 14:50 - Federated Learning Example

14:50 - 15:00 - Break

15:00 - 16:00 - Federated Learning Challenges

​

2

3 of 85

AIDA & AIDA Data Hub

​

AIDA Community - medtech4health.se/aida

National collaboration arena for AI research and innovation in medical imaging diagnostics.

​

AIDA Data Hub - datahub.aida.scilifelab.se

E-infrastructure set up to support AIDA.

​

​

4 of 85

Vetenskapsrådet

Research funding agency

Government

Executive branch

Knut and Alice Wallenberg foundation

Private research funder

VINNOVA

Innovation agency

SciLifeLab

Life science research infrastructure/center

AIDA

Collaboration arena in Swedish medical imaging diagnostics AI innovation

AIDA Data Hub

e-infrastructure supporting AIDA

NBIS

Bioinformatics platform

5 of 85

AIDA mission

Bridge the gap between promising research results and actual patient benefit, through a clinic-native research agenda for innovation.

​

​

Clinical wilderness

Research sandbox

6 of 85

AIDA Community

Publicly funded collaboration arena for AI innovation in medical imaging diagnostics.

Healthcare

Academia

Industry

7 of 85

AIDA Data Hub

E-infrastructure for clinical innovation in data driven precision health.

Data services

  • Access high quality datasets
  • FAIR sharing of DOI citable datasets
  • Extract clinical data for research

Secure long term primary storage and compute

  • Advanced data usage patterns: �Collaborate, annotate, federate, ...
  • Compute for AI training

Support

  • Data sharing, ethics, legal, and policy
  • AI development & System design

8 of 85

AIDA Data Hub Team

Caroline Bivik Stadler

AIDA Arena Director

Varshith Konda

Systems development

Pontus Freyhult

IT Architect

Betul Eren

Data sharing lead

Erik Ylipää

Support lead

Emre Balsever

Systems development

Claes Lundström

AIDA Scientific Director

Joel Hedlund

AIDA Data Hub Lead

9 of 85

Federated Learning

Why? What? How?

9

10 of 85

The tenets of modern AI

10

Rich Sutton, The Bitter Lesson (2019) http://www.incompleteideas.net/IncIdeas/BitterLesson.html

​

“One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning.”

Kaplan, Jared, et al. "Scaling laws for neural language models." arXiv preprint arXiv:2001.08361 (2020).

11 of 85

Statistical Learning Primer

11

World

Model

12 of 85

Data bottlenecks contemporary AI

12

  • Researchers at OpenAI’s office in San Francisco developed a tool to transcribe YouTube videos to amass conversational text for A.I. development. Jason Henry for The New York Times

13 of 85

Distributed Machine Learning

  • Do heavy computation in parallel, on multiple machines if needed.
    • Speedup proportional to the number of machines
  • Routine, used all the time for training large neural networks
  • Most contemporary frameworks strive to make distributed training easy

13

Large dataset

Machine 1

Machine 3

Machine 2

Partition dataset

Parameter server

Local parameter updates

Aggregated (global) parameters updates

14 of 85

Clinical data

While the legality of vacuuming the internet of data to train AI might be unclear, for health data it’s not.

Gathering all the world's health data to a supercomputer will not happen.

Alternative:

  • Share models instead of data
  • Share synthetic data

​

14

15 of 85

Share models instead of data

While the data itself might not be sharable, models derived from the data could be

A federation can be formed around this setup, and training models together becomes federated learning

15

Organization 1

Private data 1

Organization 2

Private data 2

Organization 3

Private data 3

Share model with federation

Federation

16 of 85

Typically Centralized

Model is typically shared through a common server

Clients (organizations) share the local model they have trained on their private data with central server

Server aggregates local models into global model

16

Organization 1

Organization 3

Organization 2

Aggregator

Local parameter updates

Aggregated (global) parameters updates

Private data 1

Private data 2

Private data 3

17 of 85

Training is [often] synchronized

17

Aggregator

Organization 1

Private data 1

Organization 2

Private data 2

Organization 3

Private data 3

Aggregator

Starts with global model, newly initialized or result of previous round

Train on private data

Train on private data

Train on private data

Share global model with federation

Share local models with aggregator

Aggregate local models to new global model

Adapt global model on private data into local model

One round of federated learning

18 of 85

Federated algorithms (how to aggregate)

There are many ways one can aggregate the models:

Federated Average (FedAvg) is a commonly used baseline: global model is the average of the local models

Easily extends into weighted variants (e.g. set weight inversely proportional to dataset size)

Other extensions involves server-side momentum-based methods (e.g. server has exponential moving average of aggregated gradients).

​

18

Aggregator

Aggregate local models to new global model

19 of 85

How long is a round?

19

Aggregator

Organization 1

Private data 1

Organization 2

Private data 2

Organization 3

Private data 3

Aggregator

Starts with global model, newly initialized or result of previous round

Train on private data

Train on private data

Train on private data

Share global model with federation

Share local models with aggregator

Aggregate local models to new global model

Adapt global model on private data into local model

One round of federated learning

How long do the clients train on their local data?

  • One step (e.g. batch)?
  • Fixed number of steps?
  • One epoch (go through all training data once)?
  • Multiple epochs?

20 of 85

Data non-IID

It is highly likely that data is not identically distributed across sites.

Local models are likely to diverge from each other, they learn different patterns from their respective datasets

The average of these diverged local models might perform poorly

The longer we train local models before synchronization, the larger the potential divergence

​

20

Zhu, Hangyu, et al. "Federated learning on non-IID data: A survey." Neurocomputing 465 (2021): 371-390.

An analogy: The average Gaussian fit of two independent Gaussian is a poor representation of either of them

21 of 85

Synchronize often?

Issues with divergence indicate we should synchronize often!

But, each synchronization requires us to upload/download the models to/from the server

If we synchronize too often, we mostly spend time waiting on transfers!

21

Zhu, Hangyu, et al. "Federated learning on non-IID data: A survey." Neurocomputing 465 (2021): 371-390.

Aggregator

Organization 1

Private data 1

Organization 2

Private data 2

Organization 3

Private data 3

Aggregator

Starts with global model, newly initialized or result of previous round

Train on private data

Train on private data

Train on private data

Share global model with federation

Share local models with aggregator

Aggregate local models to new global model

Adapt global model on private data into local model

One round of federated learning

Communication has a cost!

22 of 85

How long is a round?

22

Aggregator

Organization 1

Private data 1

Organization 2

Private data 2

Organization 3

Private data 3

Aggregator

Starts with global model, newly initialized or result of previous round

Train on private data

Train on private data

Train on private data

Share global model with federation

Share local models with aggregator

Aggregate local models to new global model

Adapt global model on private data into local model

One round of federated learning

How long do the clients train on their local data?

  • One step (e.g. batch)?
  • Fixed number of steps?
  • One epoch (go through all training data once)?
  • Multiple epochs?

Run experiments!

23 of 85

Evaluation - client-side

23

Aggregator

Organization 1

Private data 1

Organization 2

Private data 2

Organization 3

Private data 3

Aggregator

Starts with global model, newly initialized or result of previous round

Train on private data

Train on private data

Train on private data

Share global model with federation

Share local models with aggregator

Aggregate local models to new global model

Adapt global model on private data into local model

One round of federated learning

Aggregator

Organization 1

Private data 1

Organization 2

Private data 2

Organization 3

Private data 3

Evaluation private dev data

Evaluation on private dev data

Evaluation on private dev data

Share global model with federation

Aggregate local performance metrics to global

Evaluate global model on private development data

AUCROC2

AUCROC3

AUCROC1

AUCROCglobal

24 of 85

Frameworks - ordered after github activity

  • Active
  • Low - none activity
    • IBM Federated Learning
    • PySyft
    • FATE
    • FedML (rebranded as TensorOpera AI, aiming to do a lot more than FL)

24

25 of 85

What framework to choose

Pick one of:

  • Flower Framework, popular, requires more effort in setting up infrastructure (e.g. provisioning of client credentials, setting up TLS communication). Mainly supports python.
  • NVIDIA Flare, easy to deploy and manage clients and credentials. Mainly supports python.
  • Scaleout FedN, commercial, focus on robustness and scalability. Python, C++ and Kotlin clients. Swedish.

25

26 of 85

NVIDIA Flare

  • Technological framework for federated learning
    • Server-client setup
    • Helps setting up federated network, managing participants and cryptographic credentials
    • Building blocks for writing programs which will run on the network
    • Tools to test programs, e.g. simulating federated run
    • Some policy tools, e.g. hooks for adding custom authorization of tasks

26

27 of 85

NVIDIA Flare - Basic operation

27

Python program

28 of 85

How to determine what software to run?

  • While Flare can distribute the software to run with the task, no participant should allow unvetted code to run with their data
  • Instead, the code to run should be approved and pre-installed on the client, and when the client is started it’s done inside a environment where that code exists.
  • The server only tells it when to run the pre-installed code

28

Python program

29 of 85

What to run – implications

  • Software for each experiment has to be approved and pre-installed by all participants
  • Each experiment to be run requires preparations all over the federation
  • No one participant can decide what the others should run – automated experimentation is hard

29

Python program

30 of 85

Execution setup

30

Flare server

Flare client 1

Private data 1

Flare client 2

Private data 2

Flare client 3

Private data 3

Prepares machine by setting up software and preparing data

Asks machine to run pre-installed program - gives out weights

Asks machine to run pre-installed program - gives out weights

Asks machine to run pre-installed program - gives out weights

Program reports updated weights back

Program reports updated weights back

Program reports updated weights back

Prepares machine by setting up software and preparing data

Prepares machine by setting up software and preparing data

Org 1

Org 2

Org 3

Aggregates weights

31 of 85

Questions?

31

32 of 85

Break

32

33 of 85

Federated Learning Example

33

Peter Lundberg, Erik Ylipää, Oskar Jerdhaf, Arne Jönsson, Emma Eneling

Community Council Östergötland, Linköping University, Center for Medical Imaging and Visualization

34 of 85

Federated health

34

35 of 85

Magnetic Resonance Imaging

35

Fantastic method of looking inside the body

Extremely strong magnet

36 of 85

Medical Implants

36

Deep Brain Stimulation (DBS):

‘pacemaker for the brain’

[P. Lundberg]

37 of 85

Medical implants and MRI can be a harmful combination

37

[P. Lundberg]

The Magnet is always on

38 of 85

MRI Safety concerns

38

[Lundberg P et al, Eur Radiol (2024)]

39 of 85

MRI Safety concerns

39

[Lundberg P et al, Eur Radiol (2024)]

40 of 85

The current clinical workflow

40

41 of 85

Extend the current flow – find implant terms

41

42 of 85

Available data

Clinical notes

  • ~102000 patients
  • Written with efficiency to communicate with another human expert
  • They can be ungrammatical, contain typos, and idiosyncratic abbreviations, which can be specific to a given clinic.
  • Highly sensitive data
  • No labeled notes (notes where the implant is explicitly highlighted)

​

Implant term list

  • Rough list of ~1200 implant terms extracted from database for MRI safety
  • Includes some spelling variations
  • Scope is limited, doesn’t include all implants

​

42

43 of 85

Keyword search challenges

43

We only find the terms we know exactly, what about terms which means the same thing?

44 of 85

Automatically extract information from “natural language”

44

Natural language

Formal language

Text

summarization

Document classification

Automatic translation

Named Entity Recognition

45 of 85

What makes natural language hard?

45

“I shot an elephant in my pajamas”

No, he doesn’t have enough money to buy the car

There’s no car he doesn’t have enough money to buy

The car is beyond his budget

Superficial similarity does not mean similar meaning

Even complete superficial similarity can have different meanings

46 of 85

Two main approaches: Statistical and rule based

46

Rule based (grammars, structure)

Statistical (conditional probabilities, data, search)

David Bamman, Info 159/259, Lecture 12 - Syntax, https://people.ischool.berkeley.edu/~dbamman/nlp22.html

47 of 85

Statistical Learning Primer

47

World

Model

48 of 85

Our approach - semantic vector search

48

Word vectors - each word is encoded as a vector in a space, where different directions in space encode different semantics

Similar objects represented as vectors can be found by looking at how close they are in the semantic space

49 of 85

Semantic search with (BERT-style) LLM

49

  • We can find new glossary terms from the clinical notes by searching for words which are similar to glossary words or are used in a similar context.
  • Similarity can be assessed by measuring the distance between the embeddings of different terms.
  • XLM-Roberta will be used for representations

​

XLM-Roberta

50 of 85

Langue modelling fine tuning

  • Existing model (XLM-RoBERTa) is fine-tuned on de-identified clinical notes

XLM-RoBERTa is a multilingual model trained on 2.5TB or text from 100 languages.

  • Trained using “Masked” Language Modelling, model fills in masked words to learn word vectors based on surrounding context
  • Clinical notes are divided into context windows (512 tokens, ~400 words) the model can handle
  • Fine-tuned model is used to create “contextual” word vectors for all words in the training data set, marking them as being in the implant list or not

50

51 of 85

Experiment set-ups

2 main branches:

    • federated model trained on all data
    • local model trained only on the sites own data (one such model per site)

51

Flare server

Flare client 1

Private data 1

Flare client 2

Private data 2

Flare client 3

Private data 3

Prepares machine by setting up software and preparing data

Asks machine to run pre-installed program - gives out weights

Asks machine to run pre-installed program - gives out weights

Asks machine to run pre-installed program - gives out weights

Program reports updated weights back

Program reports updated weights back

Program reports updated weights back

Prepares machine by setting up software and preparing data

Prepares machine by setting up software and preparing data

Org 1

Org 2

Org 3

Aggregates weights

Federated (global) model

Flare client 1

Private data 1

Org 1

Flare client 2

Private data 2

Org 2

Flare client 3

Private data 3

Org 3

Local model org 1

Local model org 2

Local model org 3

52 of 85

Example evaluation (implant detection)

52

52

XLM-Roberta

XLM-Roberta

Federated (global) model

Local model org 1

Compare how well implants are found between global and local model

53 of 85

Status

Technical foundation in place, 4 sites has run a federated training on toy data (public domain texts).

4 Sites has performed local training: done masked language modelling on their own clinical datasets.

Technically, we just need to change what texts the clients point to.

  • There are other challenges, more on this after the next break

53

Flare server

Flare client 1

Private data 1

Flare client 2

Private data 2

Flare client 3

Private data 3

Prepares machine by setting up software and preparing data

Asks machine to run pre-installed program - gives out weights

Asks machine to run pre-installed program - gives out weights

Asks machine to run pre-installed program - gives out weights

Program reports updated weights back

Program reports updated weights back

Program reports updated weights back

Prepares machine by setting up software and preparing data

Prepares machine by setting up software and preparing data

Org 1

Org 2

Org 3

Aggregates weights

54 of 85

Break

54

55 of 85

Federated Learning Challenges

Distributed machine learning is easy – Federated Learning is hard

55

56 of 85

Is it safe to share the model?

Some attack vectors for shared models:

  • Inversion attacks: reconstruct fragments of training data
  • Inference attacks: infer whether some sample was part of training data, statistics of the training data or attributes of samples
  • Poisoning attacks: malicious actor injects model updates which makes it misbehave under specific circumstances (e.g. gross misprediction based on specific inputs)

56

Carlini, Nicholas, et al. "Extracting Training Data from Large Language Models." USENIX Security Symposium. Vol. 6. 2021.

57 of 85

Add noise to updates

We can disturb the model updates, by clipping gradients and adding noise

This makes inversion and inference attacks harder, but trades off performance. Differential Privacy gives the theoretical foundation for this and can guide you in how much noise is needed to bound the probability of revealing a data point

57

Aggregator

Organization 1

Private data 1

Organization 2

Private data 2

Organization 3

Private data 3

Aggregator

Starts with global model, newly initialized or result of previous round

Train on private data

Train on private data

Train on private data

Share global model with federation

Share local models with aggregator

Aggregate local models to new global model

Adapt global model on private data into local model

One round of federated learning

Add noise to model updates

Add noise to model updates

​

Add noise to model updates

​

58 of 85

Limit model capacity

Another mitigation is to limit the number of parameters the model has to learn the private data.

If the model is fine-tuned from a pre-trained foundation model (as is often the case), we can limit the number of parameters we train. Low Rank Adaptation (LoRA) is a common technique for this.

58

https://oneminutenlp.com/p/low-rank-adaptation

​

59 of 85

Other LoRA benefits

Apart from making inference and inversion attacks more difficult, LoRA also dramatically reduce the number of parameters (e.g. from 125M to 200k), drastically reducing communication overhead.

It also tends to have a regularizing effect (much fewer parameters to overfit)

59

https://oneminutenlp.com/p/low-rank-adaptation

​

60 of 85

Secure aggregation

60

61 of 85

Federated Machine Learning - just a special case of distributed machine learning?

  • Technically very similar to distributed machine learning
    • Difference is that data is private
  • Some additional complications
    • Computation (incl. result) must not leak private data (we’ve covered this)

61

61

Machine 1

Machine 3

Machine 2

Parameter server

Local parameter updates

Aggregated (global) parameters updates

Private data 1

Private data 2

Private data 3

Large dataset

Machine 1

Machine 3

Machine 2

Partition dataset

Parameter server

Local parameter updates

Aggregated (global) parameters updates

62 of 85

Distributed Machine Learning - Control

62

Large dataset

Machine 1

Machine 3

Machine 2

Partition dataset

Parameter server

Local parameter updates

Aggregated (global) parameters updates

Single organization

Physical

Software

Communication

Data

System

Access

Principal investigator

Security expert

Machine learning expert

Domain expert

Network expert

Policy expert

System administrator

Data controller

What machines are used, where are they located

How are the machines configured

What software is being run. What model is trained, how is the training defined, what hyper parameters are used.

What fabric is used for communication. How do machines communicate with each other. What machines can communicate with each other.

Who can access the machine and data. Who can configure the system.

What does the data contain. How much data is there. What data will be used for what (e.g. train vs. test)

63 of 85

Federated Machine Learning - Coordination

63

Physical

Software

Communication

Data

System

Access

Machine 1

Machine 3

Machine 2

Parameter server

Local parameter updates

Aggregated (global) parameters updates

Private data 1

Private data 2

Private data 3

Principal investigator

Security expert

Machine learning expert

Domain expert

Network expert

Policy expert

System administrator

Data controller

Organization 1

Principal investigator

Security expert

Machine learning expert

Domain expert

Network expert

Policy expert

System administrator

Data controller

Organization 2

Principal investigator

Security expert

Machine learning expert

Domain expert

Network expert

Policy expert

System administrator

Data controller

Organization 3

Controlled by each participant - minimal coordination needed

Controlled by each participant - some coordination needed

Controlled by the federation - much coordination needed

Controlled by the federation - much coordination needed

Controlled by the participant - some coordination needed

Controlled by each participant - some coordination needed

What machines are used, where are they located

How are the machines configured

What software is being run. What model is trained, how is the training defined, what hyper parameters are used.

What fabric is used for communication. How do machines communicate with each other. What machines can communicate with each other.

Who can access the machine and data. Who can configure the system.

What does the data contain. How much data is there. What data will be used for what (e.g. train vs. test)

64 of 85

NVIDIA Flare (and other frameworks)

64

Physical

Software

Communication

Data

System

Access

Controlled by each participant - minimal coordination needed

Controlled by each participant - some coordination needed

Controlled by the federation - much coordination needed

Controlled by the federation - much coordination needed

Controlled by the participant - some coordination needed

Controlled by each participant - some coordination needed

What machines are used, where are they located

How are the machines configured

What software is being run. What model is trained, how is the training defined, what hyper parameters are used.

What fabric is used for communication. How do machines communicate with each other. What machines can communicate with each other.

Who can access the machine and data. Who can configure the system.

What does the data contain. How much data is there. What data will be used for what (e.g. train vs. test)

  • Solves mainly technical problems, some coordination
  • Essentially a distributed computing framework with cryptographically secure communication
    • Helps with configuring cryptographic credentials and manage the federation
    • Communication and aggregation infrastructure

65 of 85

Technologists are used to automate everything

In a single organization setup, the ML researcher can do most things at will; data exploration, data pre-processing, setting up experiments and evaluating experiments

Automation across a federation is much harder, most steps need coordination!

65

Large dataset

Machine 1

Machine 3

Machine 2

Partition dataset

Parameter server

Local parameter updates

Aggregated (global) parameters updates

66 of 85

Repeat: What to run – implications

  • Software for each experiment has to be approved and pre-installed by all participants
  • Each experiment to be run requires preparations all over the federation
  • No one participant can decide what the others should run – automated experimentation is hard

66

Python program

67 of 85

What can you automate centrally?

  • Centrally do data pre-processing?
  • Centrally filter data to use?
  • Centrally make dataset splits?
  • Centrally prepare software environments?
  • Automatically make clients connect to the federation?
  • Automatically push software updates to clients?
  • Automatically start jobs?

67

68 of 85

What can you automate centrally?

  • Centrally do data pre-processing?
  • Centrally filter data to use?
  • Centrally make dataset splits?
  • Centrally prepare software environments?
  • Automatically make clients connect to the federation?
  • Automatically push software updates to clients?
  • Automatically start jobs?

68

You don’t know how the data is formatted, what idiosyncrasies it has and what needs to cleaned up.

69 of 85

What can you automate centrally?

  • Centrally do data pre-processing?
  • Centrally filter data to use?
  • Centrally make dataset splits?
  • Centrally prepare software environments?
  • Automatically make clients connect to the federation?
  • Automatically push software updates to clients?
  • Automatically start jobs?

69

Even if you knew how to filter, you can’t be allowed to (centrally limiting to small subsets would be a privacy risk).

70 of 85

What can you automate centrally?

  • Centrally do data pre-processing?
  • Centrally filter data to use?
  • Centrally make dataset splits?
  • Centrally prepare software environments?
  • Automatically make clients connect to the federation?
  • Automatically push software updates to clients?
  • Automatically start jobs?

70

Similar to filtering, but additionally you lack domain knowledge about the specific sites biases to make good selection for a test set.

71 of 85

What can you automate centrally?

  • Centrally do data pre-processing?
  • Centrally filter data to use?
  • Centrally make dataset splits?
  • Centrally prepare software environments?
  • Automatically make clients connect to the federation?
  • Automatically push software updates to clients?
  • Automatically start jobs?

71

Each site must be in control of what software will run. If centrally prepared, the sites still need to thoroughly inspect the environment. Probably better to inspect the build files (dockerfile, python requirements) and build your own.

72 of 85

What can you automate centrally?

  • Centrally do data pre-processing?
  • Centrally filter data to use?
  • Centrally make dataset splits?
  • Centrally prepare software environments?
  • Automatically make clients connect to the federation?
  • Automatically push software updates to clients?
  • Automatically start jobs?

72

Most setups disallow the central server to establish connections into the clients (only allowing the reverse). Some separate communication channel for initiating connections could be established with additional complexity. Often requires manual intervention (recover from crashes, network errors, power outage)

73 of 85

What can you automate centrally?

  • Centrally do data pre-processing?
  • Centrally filter data to use?
  • Centrally make dataset splits?
  • Centrally prepare software environments?
  • Automatically make clients connect to the federation?
  • Automatically push software updates to clients?
  • Automatically start jobs?

73

No

74 of 85

What can you automate centrally?

  • Centrally do data pre-processing?
  • Centrally filter data to use?
  • Centrally make dataset splits?
  • Centrally prepare software environments?
  • Automatically make clients connect to the federation?
  • Automatically push software updates to clients?
  • Automatically start jobs?

74

Probably ok, but needs to be pre-approved

75 of 85

What happens in case of drop-out?

Depends on task

In cross-silo training, losing a client is often cause to stop that training

Often

​

75

76 of 85

Federated learning is personnel intense

Many competences need to be replicated in each organisation.

76

Principal investigator

Security expert

Machine learning expert

Domain expert

Network expert

Policy expert

System administrator

Data controller

Organization 2

Principal investigator

Security expert

Machine learning expert

Domain expert

Network expert

Policy expert

System administrator

Data controller

Organization 1

Principal investigator

Security expert

Machine learning expert

Domain expert

Network expert

Policy expert

System administrator

Data controller

Organization 3

Software

Communication

Data

System

Access

77 of 85

Federated learning is more about Project Management than distributed machine learning

77

Principal investigator

Security expert

Machine learning expert

Domain expert

Network expert

Policy expert

System administrator

Data controller

Organization 2

Principal investigator

Security expert

Machine learning expert

Domain expert

Network expert

Policy expert

System administrator

Data controller

Organization 1

Principal investigator

Security expert

Machine learning expert

Domain expert

Network expert

Policy expert

System administrator

Data controller

Organization 3

Software

Communication

Data

System

Access

78 of 85

Questions

78

79 of 85

Thank you!

79

80 of 85

CCÖ Local+Federated training: Experiences

80

81 of 85

Software environments

Since we can’t just “push” software installation to clients, everything needs to be manually replicated at each site.

Differences in systems at sites makes software management more of a challenge.

​

81

82 of 85

Data preprocessing

Data has to be handles out-of-sight, difficult to look at concrete data examples from different sites.

For MLM, relatively easy, we just need text. For labeled tasks or multimodal learning this would be much more of a challenge.

82

83 of 85

Secure compute environment

Difficult to work with the local environment. Firewalls (a necessity) makes installing software more difficult.

Changes made to local software can’t be synchronized back to external repo.

Limited remote desktop, software can’t be directly uploaded.

83

84 of 85

Hardware limitations

Secure compute environments might have limited hardware (procured for the project), limitations in storage and compute compared to High Performance Compute (HPC) environments.

​

84

85 of 85

Coordination

Synchronizing between busy people is hard!

85

Principal investigator

Security expert

Machine learning expert

Domain expert

Network expert

Policy expert

System administrator

Data controller

Organization 2

Principal investigator

Security expert

Machine learning expert

Domain expert

Network expert

Policy expert

System administrator

Data controller

Organization 1

Principal investigator

Security expert

Machine learning expert

Domain expert

Network expert

Policy expert

System administrator

Data controller

Organization 3

Software

Communication

Data

System

Access