SAN FRANCISCO – Security as a key component of responsible artificial intelligence (AI) is a complex topic. On one hand, good security is essential for any AI system that uses personal data because it helps to protect it from unauthorized access, use or disclosure. On the other hand, security measures can sometimes conflict with principles of responsible AI, such as transparency and explainability.

For example, encrypting data can make it more difficult to understand how an AI system works, which can make it more difficult to hold the system accountable for its decisions.

[ Follow SDXCentral’s complete RSA Conference 2023 coverage ]

Ultimately, the goal of responsible AI is to develop and deploy AI models that are both secure and fair – meaning with the least amount of bias possible. This requires a careful balancing of different principles and considerations. There are few single “right” answers, and the best approach will vary depending on the specific use case.

A panel of pros from various sectors of the security world, who has each have organized and managed so-called red teams at their employers, discussed these conflicting issues today at RSA Conference 2023.

With the advent of freely available ChatGPT, Bard, Bing and other AI bots, things have changed universally. “We live in a totally different world that's 6-months-old, ”Rumman Chowdhury of Bias Buccaneers, said.

One of the keys to this for data scientists and developers is getting the AI models right – and up to date – for use in the real world. Security then can step in and protect the models from intruders intent upon creating mayhem in the networks. For example, as of this month, ChatGPT's mindstore did not yet include data from 2022 or 2023.

Some of the key challenges to ensuring that security is a key part of responsible AI include the following:

• Data privacy: AI systems often rely on large amounts of personal data (PII) to train and operate. This data must be protected from unauthorized access, use, or disclosure.
• Bias: AI systems can be biased, which can lead to unfair or discriminatory outcomes. Security measures can sometimes exacerbate bias, as they can make it more difficult to understand how an AI system works.
• Explainability: AI systems should be explainable, so that users can understand how they work and make informed decisions about their use. Security measures can sometimes make it more difficult to explain how an AI system works. The challenge is to get these two conflicting items to understand each other and enable collaboration across legacy boundaries.

Data privacy is a key result of excellent security. One of the main things AI software makers have to be aware of is to make sure their large data models using private information or regulated data are tightly secured, no matter how much of it is available publicly, Vijay Bolina, red team director at Deep Mind, said.

“You have to think about this from a preventative mindset: How can I do this across all my datasets and in my data planning? This should be foundational for everybody,” Bolina said.

Daniel Rohrer, red team director at Nvidia, said that AI developers need to have a common codex for security to work well in AI-reponsible applications. “Good engineering in (security software) is probably the most important aspect of all this. Bridging the gaps between what engineering, legal and users need in the code itself is important.”