AI Input Caching and Attack Filtering for Embedded Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI systems in IoT and embedded devices perform repetitive calculations even for stable inputs, leading to increased energy consumption and slowed performance due to the assumption that inputs are always new, which is inefficient for low-frequency signal changes.
Innovation Solution
Implementing a caching mechanism in AI systems to store fixed input and output pairs, allowing for immediate output provision without recalculating AI models when identical inputs are detected, and incorporating a defense module to identify and prevent adversarial attacks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If AI models are deployed in pipeline/serial fashion for embedded/IoT devices and assume all data is new, then the system can process real-time data, but it results in repetitive calculations and increased energy consumption
Solution Approach 1:
The patent implements a caching mechanism that stores previously processed input-output pairs before they are needed again. When new input arrives, the system first checks the cache for matching inputs and returns cached outputs without reprocessing, thereby performing the useful action in advance and avoiding repetitive calculations.
Solution Approach 2:
The patent creates simplified copies of previous processing results in the form of cached input-output pairs. Instead of re-running complex AI model calculations, the system uses these cached copies to quickly retrieve outputs for identical or similar inputs, significantly reducing computational overhead and energy consumption.
2Device complexity
If AI models are deployed in pipeline/serial fashion and assume all data is new, then the system can maintain simple architecture, but it slows down the overall system performance
Solution Approach 1:
The patent segments the AI processing system into distinct components: a caching module that handles quick lookups of previous inputs, and the main AI model pipeline that processes only new or different inputs. This segmentation allows the system to bypass repetitive processing for stable inputs while maintaining the integrity of the original AI model architecture.
Solution Approach 2:
The patent introduces a cache as an intermediary layer between the input data stream and the AI model pipeline. This intermediary component compares new inputs with cached entries and determines whether AI model processing is necessary, thereby improving throughput without requiring fundamental changes to the underlying AI model architecture.
3Reliability
If the AI system processes all inputs through the model, then it can ensure security against adversarial attacks, but it increases computational overhead and energy consumption
Solution Approach 1:
The patent applies preliminary filtering through the caching mechanism before inputs reach the AI model and defense model. By identifying and returning cached outputs for stable, non-adversarial inputs, the system prevents unnecessary processing that would consume energy without adding security value, while still maintaining security checks for genuinely new or potentially adversarial inputs.
Data Source
AI summary
An Artificial Intelligence (AI) system for processing of an input and a method thereof is disclosed. The AI system includes an input interface, an output interface, a defense module, an AI module, and at least a cache processing unit. The defense module is adapted to receive input from the at least one user and identify an attack vector amongst the input. The cache processing unit is configured to store characteristics of the received inputs. The characteristics of the received inputs include at least an identification information of the attack vector. The cache processing unit restricts the AI module from processing the input on recognition of attack vector in the input. The AI module processes input data in dependance of communication received from the cache processing unit and the defense module.