The Data Exchange Podcast: Ram Shankar on tools and best practices for building secure and trustworthy ML systems.
In this episode of the Data Exchange I speak with Ram Shankar, a Berkman Klein Center affiliate, and a researcher and engineer who works at the intersection of Machine Learning and Security. This episode is focused on the current state of tools and techniques for securing machine learning applications.
As companies deploy more and more machine learning models, they are beginning to develop tools and best practices for productionizing machine learning. There is a growing community around MLOps, a set of tools and practices that draws inspiration from established practices in software engineering (including CI/CD, QA/testing).
On the other hand, there is a growing awareness that AI and machine learning models need to be used responsibly and ethically (Responsible AI). To that end companies are trying to ensure that their ML models are fair, safe and reliable, transparent, secure, and protect privacy. Ram and I discussed the key challenges facing companies beginning their journey towards responsible AI, and we focused specifically on security and privacy.
He also provided concrete tips and recommendations on how to get started securing machine learning systems. At a minimum, as companies invest more in machine learning they will need to protect their IP against adversarial attacks, including model extraction and model replication:
We constantly see even academic researchers try to topple commercial machine learning systems. … There was a paper this year … that was able to replicate ML models with 99% accuracy, with less than 400 queries. And they did this across many ML services. … Think about it if you are a business decision maker and you are investing … to gain a competitive edge, and here comes an adversary who can replicate your ML model with a few hundred queries with 99% accuracy. You are all of a sudden losing your competitive edge.
Subscribe to our Newsletter:
We also publish a popular newsletter where we share highlights from recent episodes, trends in AI / machine learning / data, and a collection of recommendations.
Related content and resources:
- A video version of this conversation is available on our YouTube channel.
- Here are a couple of resources Ram mentioned in our conversation: Trustworthy ML Initiative and The Center for Trustworthy Machine Learning. I also want to highlight the mc2 (Multiparty Collaboration + Coopetition) project from RISELab.
- Download the 2020 NLP Survey Report and learn how companies are using and implementing natural language technologies.
- Marco Ribeiro: “Testing Natural Language Models”
- Krishna Gade: “What businesses need to know about model explainability”
- Xiyin Zhou: “Detecting Fake News”
- Weifeng Zhong: “Using machine learning to detect shifts in government policy”
- Alan Nichol: “Best practices for building conversational AI applications”
Register to join live or watch on-demand.
[Image from pxhere.]