This is a guest post by Sanjay Kottaram, Chief Architect and Director of Architecture at CognitiveScale.

 

WHAT IS TRUSTED AI?

Spending on Artificial Intelligence (AI) is expected to increase more than two and a half times from $37.5 billion in 2019 to $97.9 billion in 2023, according to IDC forecasts, reflecting the enormous potential collective benefits of AI. Yet broad adoption of AI systems will not come from the promised benefits alone but will also require us to trust these dynamically evolving digital systems.

Trust is the foundation of digital systems. Without trust, artificial intelligence and machine learning systems cannot deliver on their prospective value. To trust an AI system, humans must have confidence in its decisions. However, AI-based automated decisioning systems learn and evolve over time, are suspect to data quality issues, and contain many hidden decision processing layers which can make auditability and traceability challenging.

This black-box nature of AI algorithms is creating real business and societal risks from the unknown and unintended consequences of “rogue AI.” This creates the possibility of disastrous repercussions which could include the loss of human life if an AI medical algorithm goes wrong, or the compromise of national security if an adversary feeds disinformation to a military AI system.

Additional, but nonetheless significant, consequences can arise including reputational damage, revenue losses, regulatory backlash, criminal investigation, and diminished public trust. These all lead us to the need for “Trusted AI.” Learn more here.

Why do we need trusted AI?

A system built using a Trusted AI framework will help answer these questions:

  1. Why did the system predict what it predicted? (Explanation)
  2. If a person got an unfavorable outcome from the model, what can they do to change that? (Counterfactuals)
  3. Has the AI system or model been unfair to a particular group? (Fairness/Bias)
  4. How easily can the model be fooled? (Robustness)
  5. How can the system provide results relevant to different business and IT stakeholders (Risk Executives, LOB Owners, Data Scientists/IT Implementers)

Example: If a user was denied a loan by a machine learning model, an example counterfactual explanation could be: “Had your income been $5,000 greater per year and your credit score been 30 points higher, your loan would be approved.” See Image 1.

Image 1: An illustrative example of a loan application process, with Cortex Certifai producing counterfactual explanations for the AI system/model decision vs. a black box approach.

How do you measure and manage AI trust?

To trust an AI system, we must have confidence in its decisions. We need to know that a decision is reliable and fair, that it is transparent, and that it cannot be tampered with. However, automated decisioning data and models today are primarily black boxes that often function in oblique, invisible ways for both developers as well as for consumers and regulators.

To help organizations take the first step towards building and maintaining a trustworthy and responsible AI solution, AI Global and CognitiveScale have identified six different types of business risks from automated decisioning systems that need to be managed:

Bias and Fairness

Trusted AI systems are designed to ensure that the data and models being used are representative of the real world and the AI models are free of algorithmic biases. This can help mitigate skewed decision-making and reasoning that results in errors and unintended consequences.

Explainability

AI systems built using Trusted AI principles and software understand stakeholder concerns for decision interpretability and provide business process, algorithmic, and operational transparency. This can help human users understand and trust decisions.

Robustness

As with other technologies, cyber-attacks can penetrate and fool AI systems. Trusted AI systems should provide the ability to detect and provide protection against adversarial assaults, while augmenting the understanding of how issues with data quality can impact system performance.

Data Quality

Data is the fuel that powers AI. AI systems built using Trusted AI principles are designed to ensure user visibility around data drifts, data poisoning, and check data validity and fit, while confirming the needed legal justifications to use and process the data.

Compliance

Trusted AI systems employ a holistic design, implementation, and governance model that helps ensure that AI systems operate within the boundaries of local, national and industry regulations, and are built and controlled in a compliant and auditable manner.

Performance

Performance metrics are a measure of the quality of the model's predictions. The usual performance measures for evaluating a model are accuracy, sensitivity or recall, specificity, precision, KS statistic and Area under the curve (AUC).

The Cortex Certifai Approach

Cortex Certifai can help businesses automate the management of AI business risk by addressing requirements for Trusted AI systems as mentioned above and answering pressing questions, such as:

How did the AI model predict what it predicted? ​

  • Has the model been unfair to a particular group?​
  • Is the model compliant with industry regulations? ​
  • How easily can the model be fooled?

To trust a digital system, we must have confidence in its decisions. However, AI systems learn and evolve over time, and contain many hidden decision processing layers, making auditability and explainability challenging.

Explainability, fairness, robustness, and compliance are what we consider the anchors of trustworthy AI. Leaders may soon be required to explain how their automated decision making systems are making inferences and decisions or risk diminished public trust, revenue losses, and regulatory backlash and fines.

Cortex Certifai can help take the guesswork out of understanding risks introduced by automated decisions powered by predictive models, machine learning, and other technologies. Certifai automatically scans data and black-box models to learn how to explain decisions, uncover hidden bias, and probe for weaknesses in the data.

Cortex Certifai provides this visibility as a service so an organization can gain confidence in deploying AI solutions to production.

What does Cortex Certifai need from the underlying platform?

Cortex Certifai introduces a new type of workload. This is used during development in conjunction with a variety of IDE tools employed by data scientists. These tools may include JupyterLab, Jupyter Notebook, AWS Sagemaker, Azure ML Workbench, or locally used IDEs like PyCharm and others. Once the local development is complete, a small subset of data train/test/scan experiments and evaluations need to be executed.

These activities require a platform that can elastically scale based on demand. The platform must schedule and manage a large number of long running jobs, and provide the ability to reserve compute capability for critical jobs. It must provide isolation between teams and be able to run the same scan, on-premises and remotely, on any cloud. Cortex Certifai also needs an underlying platform that provides security and audit controls to allow multiple teams to collaborate on the same cluster.

Red Hat OpenShift and CognitiveScale Cortex Certifai: A powerful AI collaboration

Red Hat OpenShift is the industry’s most comprehensive container and Kubernetes hybrid cloud platform. It provides the agility, flexibility, hybrid- and multi-cloud portability, scalability, and self-service capabilities that data scientists and application developers need to build, run, and lifecycle manage AI-models and AI-powered intelligent applications. This is possible by leveraging Kubernetes Operators, integrating DevOps capabilities, and integrating with accelerators e.g. NVIDIA GPUs.

Kubernetes Operators codify operational knowledge and workflows to automate the install and lifecycle management of containerized applications with Kubernetes. For further details on Red Hat OpenShift Kubernetes Platform for accelerating AI/ML workflows, please visit the AI/ML on the OpenShift webpage.

Both organizations are pleased to announce that Cortex Certifai Kubernetes Operator is now Certified on Red Hat OpenShift.

Now, let’s walk through the key steps to roll out new instances of Cortex Certifai on OpenShift.

Image 2: Cortex Certifai Operator can be found in the AI/Machine Learning category of OpenShift OperatorHub.

Image 3: Cortex Certifai can be installed via the OpenShift UI or alternatively using OpenShift command line tool `oc`. Make sure the cluster has the proper pull credentials.

Once the Cortex Certifai Operator has been successfully installed, users with access to the Developer Catalog will have the ability to launch a Cortex Certifai instance using the running operator (Images 4 and 5).

Image 4: On the Create Operator Subscription page, select the namespace for this installation from the dropdown (Namespaces are created by system admins as a prerequisite step). The other options are set to defaults as follows:

  • Update Channel = "stable"
  • Approval Strategy = "automatic"

Click Subscribe.

Image 5: The Installed Operators page opens with a list view of operators you've installed. Click the name of the Cortex Certifai operator you just added.

A page opens that displays an overview of your operator and four other tabs:

  • YAML
  • Subscription
  • Events
  • Cortex Certifai Operator

Image 6: Create a new Instance of “Cortex Certifai” from the “Cortex Certifai Operator” tab

Image 7: Cortex Certifai create instance spec, parameters that need to be adjusted

Based on where Red Hat OpenShift is deployed (AWS/Azure/Datacenter) these parameters will need to be adjusted.

Edit these parameters irrespective of cloud or datacenter

  • spec/console/replicas
  • spec/console/route-type
  • spec/reference-model/enabled
  • spec/scan-dir
  • spec/console/s3/bucket-name
  • spec/console/s3/access-key
  • spec/console/s3/secret-key
  • spec/console/s3/endpoint
  • spec/console/azure/account-name
  • spec/console/azure/account-key
  • spec/console/azure/sas-token

Edit these parameters, If the selection is AWS Cloud or Data Center that uses s3 compliant storage like s3, NooBaa or Ceph.

Image 8: Cortex Certifai parameters if deployed in AWS or Datacenter and a compliant s3 storage is used

Edit these parameters, if the selection is Azure Cloud:

Image 9: Cortex Certifai parameter if deployed in Azure Cloud with Azure Blob Storage

Here’s my sample config that was used to create Cortex Certifai instance:

Image 10: Sample Config for Cortex Certifai on AWS using s3

Note: Parameter spec/console/route-type is set to oauth, with this we can leverage Red Hat OpenShift SSO for login to Cortex Certifai console.

After the operator is installed the following page will show up, the operator will create a set of deployments, pods, services and routes as shown in the following images. Please wait a few minutes for the Operator to pull down all the images and configure Cortex Certifai.

Image 11: Cortex Certifai Operator Installed

Image 12: Cortex Certifai Deployments

Image 13: List of Pods installed

Image 14: List of services installed

Building Trusted AI Application on Red Hat OpenShift, Enterprise Kubernetes Platform (1)-20

Image 15: Routes tab will provide the URL to access Cortex Certifai console. It was created during the installation of the Operator and spec/console/route-type “oauth” was selected. This added route to enable SSO with RHOS login.

Image 16: Login

Building Trusted AI Application on Red Hat OpenShift, Enterprise Kubernetes Platform (1)-22

Image 17: Login using OpenShift console user

The following UseCase view will be displayed.

Sample reference Use Cases have been packaged with the toolkit so open the help menu and follow the instruction to download the toolkit and install sample reports.

Building Trusted AI Application on Red Hat OpenShift, Enterprise Kubernetes Platform (1)-23

Image 18: Download “Certifai Toolkit”

Follow next few steps to unzip and upload the sample report to AWS s3

unzip certifai_toolkit.zip

Set the following environment variables accordingly:

ACCESS_KEY=<access key>
SECRET_KEY=<secret key>
BUCKET_HOST=<object store endpoint>
BUCKET_NAME=<bucket name>

From the directory you extracted the toolkit into, you can either use Docker, s3cmd, or your preferred tool to push the sample reports to the bucket.

Install Docker

docker run \
-v $(pwd):/s3 \
schickling/s3cmd  \
--access_key=$ACCESS_KEY   \
--secret_key=$SECRET_KEY   \
--host=$BUCKET_HOST   \
--host-bucket=$BUCKET_HOST \
--no-ssl  \
put -r /s3/examples/reports/ s3://${BUCKET_NAME}/

Or

Install S3cmd

s3cmd  \
--access_key=$ACCESS_KEY   \
--secret_key=$SECRET_KEY   \
--host=$BUCKET_HOST   \
--host-bucket=$BUCKET_HOST \
--no-ssl  \
put -r examples/reports/ s3://${BUCKET_NAME}/

Now you can view the example reports from the Certifai Console.

Building Trusted AI Application on Red Hat OpenShift, Enterprise Kubernetes Platform (1)-16

Image 19: Certifai with Sample Reports

Image 20: Option to view Details/Scan List

Image 21: Use Case Detail

Building Trusted AI Application on Red Hat OpenShift, Enterprise Kubernetes Platform (1)-18

Image 22: Use Case Scan List

Select the view results to inspect the evaluations performed by Cortex Certifai.

Image 23: Scan Details for Banking Loan Approval

Follow the documentation to understand more about Cortex Certifai: https://cognitivescale.github.io/cortex-certifai/

Here is a Reference Architecture how Trust-as-a-Service using Cortex Certifai can be enabled on Red Hat OpenShift and Open DataHub Software components.

Building Trusted AI Application on Red Hat OpenShift, Enterprise Kubernetes Platform (1)-15

Image 24: Reference Architecture for Red Hat OpenShift using Open Data Hub assets with Trust-as-a-Service enabled using Cortex Certifai

The ultimate goal is that this is not just an extended blog post, but more of a helpful tutorial for you to blend two leading technology AI platforms that move your organization forward in a way that is simple, compliant, and effective 一 and most importantly, worthy of trust from your shareholders. It’s all about working together to harness the potential that AI solutions can collectively provide.


About the author

Red Hatter since 2018, tech historian, founder of themade.org, serial non-profiteer.

Read full bio