20. Edge Computing

Edge Computing

What is Edge Computing

Edge Computing(fog computing, mesh computing)

• Access computers on the "edge", where edge is defined as somewhere close to our originating data sources
  • often involves peer-to-peer computing, although it may also involve servers which are on the edge
• Data is born at the edge: billions of smartphones & IoT devices constantly generate data
• Cloud computing refers to accessing computer resources out somewhere on the wider Internet
  • Compute has been cloud dominated

Taxonomy

• Far edge: sensor, IoT, human (data generator, ege device)
  • very limited networking
• "edge" (edge node, cloudlet)
  • local compute, storage
  • 1 hop to far edge, Internet connected
• Local cloud
  • collection of edge nodes
• Centralized cloud

 

Why Edge Computing

Centralization: High latency, Low BW <-> Dispersion: Low latency, High BW

• proximity, latency: highly responsive applications
• scalability via edge analysis: local processing of high BW sensors
• privacy enforcement
• masking cloud outages

Use case

• Video analytics

Why now?

• Networking, Computing power ⬆️
• Explosion of data at the edge

Edge Devolution

• From just a data producer to the cloud, now, it is a data consumer and producer at the same time

Edge Computing Models

• Mobile offloading: face recognition - latency & energy
• Cloud offloading: shopping cart updating - latency
• Edge data processing: localized data analytics

 

Challenges of Edge Computing

Challenges

• Reliability, naming, programming, heterogeneity, scalability
• Complexity management
• Weaker security
• Non-technical challenges
  • Who provides edge infrastructure?
  • Who funds edge business?

---

Federated Learning

Decentralized Data

• standard setting in ML considers a centralized dataset -> in the real world, data is often decentralized

Limitation of Centralized Data

• Sendinig the data is too costly
• Data may be considered too sensitive

 

Federated Learning (FL)

Federated Learning

• FL aims to collaboratively train a ML model while keeping the data decentralized
• Ideally, the final model is as good as the centralized solution
• Data is decentralized and never leaves local silos
• Central server controls training across decentralized sources
• Improves privacy and communication efficiency

Characteristics of FL

• Data is genereated locally and remains decentralized
• Each client store its own data and cannot read the data of other clients
• Clients send only weight gradients
• Data is not independently of identically distributed (non-IID)

Challengs of FL

• Algorithms
  • Traning dataa is non-IID
  • How to update local and global gradients
• Network performance
  • Learning is bottlenecked by network performance
  • Network topology changes every time
• Fault tolerance
  • Some of local devices can fail, msgs may not be delivered
• Concurrency
  • Network latency exists -> asynchronous
• Personalization
  • Local devices may use different languages
• Malicious clients
  • Weight poisoning attack