Leaks

Unlocking Human Potential: A Deep Dive into new Microsoft’s Trustworthy AI

Copyright & Fact Checker 25 October 2024

Can you trust AI? This question is at the forefront of technological advancement. Microsoft is committed to building Trustworthy AI, ensuring data privacy, security, and responsible use. Here’s a breakdown of their approach:

1. Safety Guardrails:

Real-time filters: Eliminate bias, harmful content, and misleading information.

Transparency: Understand AI’s source and reasoning behind responses.

Prompt injection attacks: Mitigate attempts to bypass safety measures.Direct attacks: User tricks to generate unwanted content.

Indirect attacks: Hidden instructions in data to manipulate AI.

2. Groundedness and Accuracy:

Context and data: Provide accurate grounding information for responses.

Groundedness detection: Identify and correct inconsistencies between data and responses.

Model Selection: Choose the right model based on your application’s needs.

3. Data Privacy:

Data control: Your data is never used for training Microsoft’s core models.

Confidential Computing: Encrypts data during use, even on GPUs, for unparalleled privacy.

Verifiable Confidentiality: Ensures every step of processing is auditable.

4. Security:

Model scanning: Identify vulnerabilities and malicious code in AI models.

Secure access control: Manage access to services, data sources, and infrastructure.

Web Query Transparency: Verify what information Microsoft Copilot uses to generate responses.

Data Loss Prevention: Prevent sensitive data leaks through access control and labeling.

Audit trails: Monitor AI app usage and access to sensitive information.

This is just the beginning! Microsoft is constantly innovating to ensure Trustworthy AI.

Want to learn more?

Microsoft Trustworthy AI: https://blogs.microsoft.com/blog/2024/09/24/microsoft-trustworthy-ai-unlocking-human-potential-starts-with-trust/

Confidential Inferencing: https://techcommunity.microsoft.com/t5/azure-confidential-computing/azure-ai-confidential-inferencing-technical-deep-dive/ba-p/4253150

References: Internet Archive, Undercode Ai & Community, Techcommunity.microsoft.com,es: AI Pioneers, Wikipedia
Image Source: OpenAI, Undercode AI DI v2Featured Image