Propagation of new typologies
By sharing the essence of the patterns investigators discover without revealing sensitive information, all the participants (or nodes in a Federated Learning network) can learn from each other and become better at their jobs.
A detective club
Think of Federated Learning as a detective club made up of banks from all over the world. Each bank (or detective) is working to spot the crafty schemes of money launderers (or criminals) in their own area. They do this by observing patterns in financial transactions, much like noticing if a criminal has a particular style of operation or a signature move.
A new typology
Now, let's say one bank in this club detects a new money laundering scheme locally. This scheme is quite ingenious, and none of the other banks have seen it before. This is equivalent to a local pattern discovered by a participant in a Federated Learning network.
Translate to code
This bank could try to describe the scheme in words to the rest of the club, but that might reveal sensitive details about its customers or operations. So, instead, it translates this scheme into a code (a model update) that doesn't reveal any specific transaction details but carries the essence of the scheme.
The bank then sends this code to a secure central server. You can imagine this server as the club's meeting place, where all banks share their updates. Here, all the individual codes from each bank are combined to create a master code (the global model update), which includes the knowledge of all banks, including the one that spotted the new scheme.
Finally, this master code is shared back with all the banks in the club. Even though they don't know the specifics, they can now recognize the new scheme if they encounter it. This way, every bank benefits from the insights of all the others, including those who detect new money laundering methods, making the whole system more effective at catching these financial criminals.
With the global code or model, each bank is now able to separately detect the new crime pattern identified by one of the network participants. No data has been shared, but insights from one participant has been distributed to all - immunizing the whole network from the new crime typology.