Learning Library

← Back to Library

Federated Learning: Model Training Without Data Sharing

Key Points

  • Federated learning flips traditional AI training by sending a shared model to each device or organization to learn locally, then returning only model updates instead of raw data.
  • Each participant (e.g., smartphones, laptops, or companies) trains a local copy of the model on its own sensitive data, preserving privacy while still contributing insights.
  • A central server aggregates these encrypted updates to continuously improve a global model, enabling collective learning without exposing individual datasets.
  • Originating at Google in 2016 amid rising privacy concerns, federated learning is illustrated by multiple companies collaboratively refining a market‑trend predictor without ever sharing proprietary sales data.

Full Transcript

# Federated Learning: Model Training Without Data Sharing **Source:** [https://www.youtube.com/watch?v=zqv1eELa7fs](https://www.youtube.com/watch?v=zqv1eELa7fs) **Duration:** 00:06:26 ## Summary - Federated learning flips traditional AI training by sending a shared model to each device or organization to learn locally, then returning only model updates instead of raw data. - Each participant (e.g., smartphones, laptops, or companies) trains a local copy of the model on its own sensitive data, preserving privacy while still contributing insights. - A central server aggregates these encrypted updates to continuously improve a global model, enabling collective learning without exposing individual datasets. - Originating at Google in 2016 amid rising privacy concerns, federated learning is illustrated by multiple companies collaboratively refining a market‑trend predictor without ever sharing proprietary sales data. ## Sections - [00:00:00](https://www.youtube.com/watch?v=zqv1eELa7fs&t=0s) **Federated Learning: Decentralized Model Training** - Federated learning trains AI models locally on devices and sends only model updates—not raw data—to a central server, preserving privacy while constructing a comprehensive global model. - [00:03:22](https://www.youtube.com/watch?v=zqv1eELa7fs&t=202s) **Federated Learning: Types and Benefits** - The passage explains how federated learning enables companies to refine a shared global model using private data updates, and outlines its three variants—horizontal, vertical, and transfer—along with their broad use cases. ## Full Transcript
0:00Let's unpack the concept of federated 0:03learning. A method for training AI 0:05models that is all about keeping your 0:07sensitive data right where it should be 0:10with you. Now AI applications like chat 0:13bots, recommendation systems, and spam 0:16filters. They're all very data hungry 0:18and they have been fed tons of examples, 0:24mountains of information which they use 0:26to learn their specific tasks to build 0:29an AI model. 0:33Now normally in machine learning we 0:35gather all of this data from different 0:36sources and bring it to one place. All 0:40of this will reside in a central server 0:44and that's where the actual training of 0:46the model takes place. Federated 0:48learning turns this process on its head. 0:50Instead of bringing the data to the 0:53model, we take the model to the data. So 0:56here's how it works. Think every device 0:59like a smartphone or a laptop or a 1:02server. It has its own local version of 1:08a model. So each of these are reporting 1:10into their own model and this model 1:13learns from the data right there on the 1:16device itself. Now after the model has 1:18learned from the local data, it sends 1:21only the model updates back to the 1:23central server, not the actual raw data. 1:27So this all goes here to the central 1:30server. And then that server aggregates 1:33all of these updates from all the 1:36devices to create what is called the 1:39global 1:41model. 1:45Now why bother with this level of 1:47decentralization? 1:49Well, this concept was first introduced 1:50by Google in 2016 at a time when global 1:53attention was focused on the use and 1:55misuse of personal data. Concerns about 1:58data privacy and security prompted the 2:01search for alternatives to traditional 2:03centralized AI training methods, giving 2:06birth to federated learning. So let's 2:09imagine a scenario involving a group of 2:11companies that want to collaborate on 2:13building a model to predict market 2:15trends, but each company has sensitive 2:19sales data they want to keep private. So 2:22each company has access to an initial 2:26baseline predictive global model. Here's 2:29our global model up here. And this 2:32resides in a central server. 2:37Now in their individual environments, 2:39each company trains the instances of the 2:42model using their own sensitive sales 2:45data. So we have the global model here 2:47and then these individual models with 2:50each company. And here is their 2:51sensitive sales data along the bottom. 2:54And they're tweaking and refining their 2:56model based on their unique data. So the 2:59companies do not share their sensitive 3:01sales data. Instead, they only share the 3:04updates they made to the model. Now 3:06these updates, they don't contain any 3:08raw sales data, but they do reflect the 3:11insights gained from the data. The model 3:14updates are then sent back to the 3:17central server. 3:20And here they're integrated into the 3:22global model. Now this iterative process 3:25continues with each company refining the 3:27model based on their private data and 3:30sharing only the model updates. Over 3:32time, this model becomes increasingly 3:34accurate at predicting market trends. 3:36Even though no company had to share 3:38their sensitive data, each company 3:39benefits from the collective 3:40intelligence of the group while 3:42maintaining their data privacy. That is 3:45the essence of federated learning. 3:47Allowing for collaborative learning from 3:49shared model updates while keeping the 3:51actual data distributed and private. Now 3:55we can think of federated learning as 3:57coming in three flavors. So there's 4:01horizontal and horizontal federated 4:04learning describes the forecasting model 4:06example we've just discussed where the 4:08data sets were all similar. In this case 4:10the similarity was this was all sales 4:12data. 4:13Now another one is called 4:17vertical federated learning. So instead 4:20of using similar data sets we're dealing 4:22with complimentary data using movie and 4:26book reviews for example to predict 4:28someone's music preferences. And then 4:31the third kind is called federated 4:34transfer learning. Here we start with a 4:37model that's already been trained to do 4:39one task and then adapt it to do 4:42something slightly different like like 4:44how a pre-trained foundation model 4:46designed to perform a task like 4:47detecting cars is trained on another 4:50data set to do something else entirely 4:52like identify cats. Now the use cases 4:55for federated learning are farreaching 4:57and impactful. Just consider the 4:58healthcare industry where federated 5:00learning allows medical institutions to 5:02collaboratively train their models on 5:04their sensitive data without sharing the 5:07actual medical records or how financial 5:09institutions can improve their fraud 5:11detection mechanisms and credit scoring 5:13systems without compromising on customer 5:15privacy. However, federated learning is 5:18not without its challenges. There is the 5:20risk of inference attacks where 5:23adversaries may try to extract 5:26information about the data from the 5:28shared model updates when we've put them 5:30up there. Now to counter this, 5:32researchers are looking into strategies 5:34like secure multi-party computation to 5:36ensure privacy by encrypting model 5:38updates or by adding a degree of noise 5:40to the data to mislead potential 5:43attackers. Other challenges include 5:45computational efficiency because we do 5:47have all of this work going on locally 5:49here and maintaining transparency in 5:52model training and creating incentives 5:54for truthful participation. But in the 5:56end, federated learning offers a 5:59promising path towards a new generation 6:01of AI applications. By addressing 6:03privacy concerns and leveraging the 6:06power of distributed computing, 6:08federated learning holds the potential 6:10to revolutionize how AI models are 6:13trained. 6:15If you have any questions, please drop 6:16us a line below. And if you want to see 6:19more videos like this in the future, 6:20please like and subscribe. Thanks for 6:23watching.