Federated learning: Intelligence versus confidentiality: can we have both?
In my previous blog I have written about AI based recommender systems and how they have changed our lives over the past decade. As I sat down to write this time, I reflected on the issues of large-scale machine learning (ML), data privacy, and federated learning (FL), an emerging trend that is a topic of research. burning for building AI models. FL is mainly a distributed ML architecture which allows the training of an algorithm on several decentralized data. Instead of being collected on a single server, data remains locked on a server or edge device, while only the algorithm moves between servers. So FL operates on the fundamental approach of bringing code to data, instead of data to code.
With growing concerns about data privacy, data security, data rights, and data access, techniques such as federated learning are gaining popularity as they provide an opportunity to reduce data risk. In this blog, I discuss some FL use cases and how NetApp can help address FL challenges associated with data protection and privacy.
Emerging Industry Use Case
Using FL and removing barriers to sharing data benefits almost every industry. In banking and financial services, FL techniques are used to optimize price and expense ratios in portfolio management. FL enables asset managers, financial advisers and robo-advisers to maintain the confidentiality of their clients regarding the components of the portfolio. FL also allows them to connect with other investment banks who can provide a fair price when buying or selling a client’s portfolio.
Financial institutions train neural network models on the server by sending encrypted model weights and bias coefficients back and forth. By dealing with financial crimes and the pressure of strict regulatory compliances like GDPR, FL systems have the potential to enhance current efforts to tackle illegal financial activities like money laundering and fraud. FL accomplishes this improvement by enabling shared machine learning without sharing data.
In health care, for example, training an AI model to identify damage to the hippocampus in an MRI of the brain is an important step in the diagnosis of patients with dementia or Alzheimer’s disease. To create robust AI algorithms, hospitals and medical research institutes often need to collaborate and pool their research knowledge, as in the case of the Personal Genome Project (PGP). However, deciding on the authorized use of data while preserving the patient’s right to privacy is a difficult task.
The EMR CXR AI (EXAM) model, a new study led by Mass General Brigham and NVIDIA, brought together 20 institutions from around the world to form a neural network. The model predicts the future oxygen needs of symptomatic patients with COVID-19, using vital signs inputs, lab data, and chest x-rays. The results of this collaborative learning effort are published in Natural medicine. The model, which is publicly available for research via the NVIDIA NGC hub, use NVIDIA Clara for training federated learning capabilities with AI-assisted annotation and transfer learning. NetAppÂ® ONTAPÂ® AI for diagnostic imaging with NVIDIA Clara provides guidance on workflows used in the development of deep learning (DL) models for medical imaging. NetApp ONTAP AI is also validated with DL platforms like TensorFlow, which offers the open source framework Federated TensorFlow for data scientists who develop AI models on decentralized data.
The two most common types of federated learning
- Multi-device federated learning is typically deployed in a single organization. It includes IoT sensors, mobile devices or peripherals that belong to users of a single organization
- Federated learning between silos Usually involves multiple organizations, such as the example of multiple hospitals and research institutes mentioned earlier in a healthcare scenario.
Design challenges with federated learning
Some possible solutions to privacy concerns are encryption (centralized) and federated learning (decentralized). However, recent research has shown that preservation of privacy in Florida Keeping data and calculations on the device is not enough to ensure privacy. This is because the template parameters exchanged between parties in an FL system always hide sensitive information, which can be exploited in privacy or security attacks like the Byzantine and data poisoning attacks in federated learning. Therefore, FL systems need effective data protection, data governance and privacy preservation across AI infrastructures that are compliant with programs such as GDPR, CCPA and LGPD. .
NetApp AI infrastructure and data privacy solutions
NetApp offers a wide variety of products and services with tools that can be used in your privacy operations, such as GDPR and CCPA compliance programs. NetApp solutions address a full spectrum of cybersecurity threats, and they do Data protection and Security assessment The right way. These solutions include:
NetApp Cloud data discovery to help you identify the personal information in your data, adopt policies, meet privacy requirements and align with data governance.
- NetApp SnapCenterÂ® technology to support backup and restore.
- NetApp FPolicy for privacy operations and policy enforcement.
NetApp StorageGRIDÂ® object store for hybrid multicloud environments.
- NetApp ONTAP data management software with unified storage and access to S3 objects for next-generation applications such as augmented and virtual reality, autonomous vehicles and cashier-less stores.
Other solutions include:
NetApp Astra â¢ control to manage, protect, and move data-rich Kubernetes (K8) workloads in the public cloud and on-premises.
- IA NetApp ONTAP powered by NVIDIA GPUs for cloud-connected AI / ML training.
- NetApp AI Inference for real-time event streaming that enables advanced computing
- NetApp AI control plane and MLRun pipeline for AI and ML model release management, A / B testing, MLOps, and serverless automation in distributed computing environments or multi-node AI and ML systems running FL aggregation algorithms like FedAvg, Scaffold, etc.
Example of a cross-silo federated learning, several organizations.
Storage, compute, local FL models owned by each organization (on-premises, hybrid cloud, edge).
Data more widely disseminated, for example between hospitals, banks, etc.
While FL’s future direction focuses on containerization and security frameworks that ensure reliable storage orchestration and alleviate concerns about data outages or federated learning failures, a robust AI infrastructure at scale is imperative.
Federated learning provides an opportunity to reduce data risk. However, it also presents new risks that have yet to be fully discovered and addressed. Every new risk can have a technological solution, but without strong data governance, data protection, and a reliable AI infrastructure, FL is unlikely to be effective. NetApp is working to create advanced tools that remove bottlenecks to help AI engineers fill some of these gaps. As a data-centric, cloud-driven software vendor, NetApp is uniquely positioned to deliver a data structure with industry-leading data management capabilities in the industry-leading cloud ecosystem. NetApp helps customers create privacy-preserving ML and AI models that enable private, secure, and transparent data analysis.
Learn more about NetApp AI Solutions.
NetApp Inc. published this content on December 17, 2021 and is solely responsible for the information it contains. Distributed by Public, unedited and unmodified, on December 18, 2021 10:09:08 AM UTC.
Public now 2021
Trends in technical analysis NETAPP, INC.
|Short term||Mid Road||Long term|
Evolution of the income statement
|Number of analysts||23|
|Last closing price||
|Average price target||
|Spread / Average target||11.5%|