Vision Processing Unit Secure AI Inference on Encrypted Video

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current techniques for machine learning and artificial intelligence inference on video frames expose protected video content, leading to undesired CPU cycle usage and proprietary software limitations, which hinders the acceleration of AI inference while maintaining media content protection.

Innovation Solution

Implementing a vision processing unit (VPU) with a neural network processor that performs AI inference on encrypted video content within a Trusted Execution Environment, using a scheduler or dispatcher for frame-level pipelining and direct memory access to secure encryption and decryption, maintaining content protection and accelerating AI processing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If end-to-end software stack is used for encryption and decryption, then content protection is maintained, but CPU cycles are consumed and AI inference is exposed

Engineering Contradiction:

Improvecontent protectionVSAvoidAI inference acceleration

Core Design Contradiction:

ReliabilityVSProductivity

Solution Approach 1:

The system segments the encryption/decryption function from the AI inference function by implementing a dedicated crypto engine separate from the CPU and AI processor. This allows encrypted video data to be processed through the AI pipeline without exposing content, while the crypto engine handles security operations independently, resolving the contradiction between content protection and AI inference acceleration.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces a crypto engine as an intermediary component between the AI processor and the video data. This intermediary handles encryption and decryption operations, enabling the AI processor to work with encrypted data without direct CPU involvement, thus maintaining content protection while accelerating AI inference.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Reliability

If CPUs are used for encryption and decryption, then content protection is maintained, but processing speed is reduced

Engineering Contradiction:

Improvecontent protectionVSAvoidprocessing speed

Core Design Contradiction:

ReliabilityVSSpeed

Solution Approach 1:

The patent replaces the mechanical CPU-based encryption/decryption system with a dedicated crypto engine implemented in hardware or optimized firmware. This substitution eliminates the need for CPU cycles to be consumed by cryptographic operations, significantly improving processing speed while maintaining content protection through the specialized crypto engine.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Reliability

If proprietary software stack is used, then content protection is maintained, but software flexibility is limited

Engineering Contradiction:

Improvecontent protectionVSAvoidsoftware flexibility

Core Design Contradiction:

ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system segments the proprietary crypto engine from the AI inference software stack. The crypto engine is implemented as a dedicated hardware or firmware component that handles security operations independently, allowing developers to write standard, flexible AI inference code without being constrained by proprietary software requirements, thus maintaining content protection while improving software flexibility.

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS11736764B2Artificial intelligence inference on protected media content in a vision processing unit

Publication Date: 2023.08.22 INTEL CORP

AI summary

Techniques related to securely providing artificial intelligence inference on protected video content in a vision processing unit are discussed. Such techniques include decrypting encrypted video via a neural network processor of the vision processing unit by providing the neural network processor direct memory access to a security engine of the vision processing unit and applying a machine learning model to the decrypted video content using the neural network processor such that a host and other components of the vision processing unit do not have access to the decrypted video content.