Privacy aware analytics at edge using Federated Learning In Cloud computing or centralized approach, data-sources or sensors collect data from environment; the data is then sent to a data center placed at a geographically different location to be processed and analyzed. This approach is not suitable for the applications required low latency and quick response time. Also, in massive machine type communications (MMTC) where thousands of sensors or IoT data sources collect and send data simultaneously, bandwidth becomes a bottleneck in data transmission. On contrary, Edge computing is a distributed computing paradigm which brings computation and data storage closer to the location where it is generated, to decrease latency and to save bandwidth. Edge analytics may not be a replacement for centralized data analytics, but both can supplement each other in delivering data insights. With the increasing demand and implementation of stringent data privacy laws and growing security concerns, the concept of Federated Learning (FL) has been introduced. In FL, edge applications can use their local device data to train an Machine Learning model required by the centralized server. The edge devices then send the model updates rather than raw data to the server for aggregation. FL can serve as an enabling technology in mobile edge networks since it enables the collaborative and decentralized learning. In this talk, we will discuss about the decentralized learning approach in federated learning, implementing federated learning in edge devices, how it solves data localization, privacy and scalability issues, federated learning at edge and fog devices with practical use cases and open problems.
Privacy aware analytics at edge using federated learning
1. Privacy aware analytics at
edge using Federated
Learning
Arindam Banerjee
Data Scientist, Ericsson
1Arindam Banerjee
2. Shaping the next decadeâŚ
⢠5G
⢠Exponential growth of
Smart devices
⢠Exponential growth of
data
⢠Stricter data privacy
and protection laws
⢠Automation and
Analytics
2Arindam Banerjee
3. 5G and IoT: Ushering in
a new era
⢠Foundation for realizing the full
potential of IoT.
⢠550 million 5G subscriptions in
2022 - Ericsson ABâs Mobility
Report.
⢠A fertile ground for innovations
and customers engagements.
⢠Network slicing. Segmenting
network:
ďenhanced Mobile Broadband
(eMBB),
ďUltra Reliable Low Latency
Communications (URLLC),
ďmassive Machine Type
3Arindam Banerjee
4. Exponential growth
of Data
⢠Billions of IoT sensors
and mobile devices at
edge.
⢠Uncertainty about network
quality.
⢠Intermittently available.
⢠From massive data to
context-aware, smarter
apps.
4Arindam Banerjee
5. Privacy and Security
⢠IoT data is growing fast but
security
remains a big hurdle.
⢠Hyper-personalization comes with
loss of privacy.
⢠Laws and regulations:
ďHealth Insurance Portability
and Accountability Act
(HIPAA),
ďGeneral Data Protection 5Arindam Banerjee
6. Problems faced in CentralizedAnalytics
Edge to cloud
data
transmission
Bottleneck -
Network
bandwidth
Security and
privacy
Infrastructure
cost
Underutilized
resources at
edge
6Arindam Banerjee
7. Federated Learning (FL)
⢠Collaborative learning
⢠Downloaded model learns from local data
⢠Compressed update is sent back to the cloud
⢠Federated averaging
⢠New model is communicated back to edge
⢠Improved local inference empowers
personalization.
7Arindam Banerjee
9. Features of Federated Learning
Edge IoT
devices have
limited
network
bandwidth.
Devices are
intermittently
available for
training.
Device may
choose not to
participate in
the training.
Massive number
of devices but
inconsistent.
Naturally
arising non-
IID partition.
Massively
parallel.
9Arindam Banerjee
10. Advantages of Federated
Learning
⢠Privacy:
ďData localization
ďData retention
⢠Less data transfer - less Network
Bandwidth
⢠Better battery life.
⢠Scalability.
⢠Less infrastructure cost.
⢠Low latency for inference.
10Arindam Banerjee
11. ⢠High convergence time.
⢠Unavailability of edge
devices.
⢠Irregular or missed
updates.
11Arindam Banerjee
12. Model Poisoning
Image Source: Clement Fung et al. â âMitigating Sybils in Federated
Learning Poisoningâ
12Arindam Banerjee