The RSA conference is one of the top industry security conferences in the world, offering four days of nonstop cutting-edge sessions, inspirational keynotes, and groundbreaking innovation.
This year it brought together thousands of industry experts and business professionals to advance awareness and understanding of critical cybersecurity issues, with over 26,000 attendees, including 600+ speakers, 400+ exhibitors, and over 400 members of the media. (https://www.rsaconference.com/events/2022-usa)
In this year’s conference, IBM presented a talk on Privacy and Compliance for AI – Open-Source Tools and Industry Perspective
The talk focused on two key aspects of adoption of AI privacy and compliance technologies: leading open-source projects, and important considerations that should be accounted for when constructing solutions for enterprise use. This was a continuation to a talk given in last year’s conference on inference attacks against machine learning models (https://www.rsaconference.com/library/Presentation/USA/2021/evasion-poisoning-extraction-and-inference-tools-to-defend-and-evaluate). This talk was more focused on possible mitigations using open-source techniques.
We started by presenting different privacy-preserving technologies for ML models, including model anonymization, differential privacy and encryption, mentioning the pros and cons of each method and its relation to the privacy/accuracy or privacy/performance tradeoff.
Then we covered aspects related to the industry perspective and relevant considerations. We stressed that any solution must be non-disruptive and be able to easily integrate with complex ML pipelines. They should also support a separation of concerns between data scientists and privacy experts. That design choices should be made that support ease of use (“one-click” solution) but are also customizable for more expert users, and provide a visualization of the tradeoffs. And finally that methods should be scalable to thousands of models and millions of records, sometimes sustained by very small teams, prioritizing automation and time-efficient algorithms.
Then we presented in more detail the open-source ai-privacy-toolkit (https://github.com/IBM/ai-privacy-toolkit), that contains tools for model anonymization and data minimization, and showed a demo of the data minimization technology in action (https://github.com/IBM/ai-privacy-toolkit/blob/main/notebooks/minimization_adult.ipynb).
223 people attended the session, with the following breakdown by role:
Abigail Goldsteen, IBM.