Cloud Native Computing Foundation KubeCon + CloudNativeCon North America 2020 - Virtual, 4 Dec 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Sponsored Keynote: Scaling Machine Learning Without Compromising Privacy - Nanda Vijaydev

Description

Don’t miss out! Join us at our upcoming event: KubeCon + CloudNativeCon Europe 2021 Virtual from May 4–7, 2021. Learn more at https://kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects.

Sponsored Keynote: Scaling Machine Learning Without Compromising Privacy - Nanda Vijaydev, Lead Data Scientist and a Distinguished Technologist, HPE

No matter what kind of machine learning (ML) applications you are building for your business, securing your end-to-end ML pipeline is essential yet complicated. ML pipelines become more powerful by distributing a complex network of stages across many environments, often with multiple users and roles. This session will cover how you can leverage open source projects such as SPIRE, OPA, and Envoy to provide a fine-grained policy overlay for your ML pipeline.

https://sched.co/eoDy

A

Hi welcome everyone. This talk is about scaling, machine learning without compromising privacy and security. Thank you cncf for giving me this opportunity and thank you audience for your attention. My name is nanda vijaydev. I lead product management at hpe as well container platform and machine learning. Ops. My background is in technology transformation at medium to large enterprises specializing in big data and machine learning operations.

A

Over the next few minutes, we will take a look at the complexities of the machine learning pipeline. What are some of the possible touch, points and areas of concern and what are some of the technologies that are being used to mitigate risks and lower the exposure in machine learning pipelines?

A

This picture is from a famous paper by google, where it talks about the technical debt and the various stages of machine learning, and most people associate machine learning with algorithms, with the ml code or the ml model itself, but there's a lot that happens before during and after you know, this may run on one or many kubernetes clusters, depending on the deployments at customers.

A

There are various actors. Now there is various personas that interact with the different stages of the pipeline, and there are services within kubernetes and ingress for users to access those services, and also there are external entity. If you look at this picture, this is a representative architecture of a large financial institution that has a pretty large footprint of machine learning activities that happen across various systems and across different groups.

A

Right. You have data pre-processing done by your data engineers, and then you have the actual machine learning process done by your data scientists and then a deployment of a model. So I'm not getting into the detail of this.

A

What's of concern for us in the next few minutes is that there are several different processes. As you can see here, this perimeter there is edge attack, surface multiple, different kubernetes clusters, probably in different trust domain, and then you have a number of different external systems that have to be accessed and for those who are familiar with security and kubernetes by default. There's really no security, that's enabled!

A

So if you do a threat model for this, you can see that the graph of how a user accesses the system and what are all the touch points they go through for various activities. What services are talking to other services, so you have authenticating and authorizing to individual service by users.

A

Then you have to establish trust between services. Tenants and you know, across sites, if needed, because of external systems that are accessed as part of this pipeline. You have an increased attack surface, especially with data access. Now you have privacy concerns here. Our path to resolution is happening in multiple phases using best of the breed technologies uh such as ldap and oidc connector. As you can see here and mtls between uh systems using a you know, jot and spiffy identities for preventing attack surface.

A

You know this is something we're paying special attention to with admission control, various part, security policies and network controls and, um for you know, extended policies such as limiting your authorized docker registry and things like that. There are oppa, regular policies for phase 2.

A

There is experimentation going on to establish more transitive identity between user services and also external resources that are outside of the community's network and for federation between on-prem and cloud trust domains. We are looking at spiffy federation.

A

Thank you and looking forward to seeing you on another talks.