On-Device AI Content Blocking With Efficient Vision Transformers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing content control technologies for children and vulnerable individuals fail to effectively prevent exposure to and creation of inappropriate content, particularly self-generated content, due to reliance on cloud connections, complex configurations, and inadequate processing power on user devices, leading to privacy risks and inaccurate content detection.
Innovation Solution
An embedded artificial intelligence (AI) system utilizing an efficient vision transformer (EVT) architecture with structural reparameterization and reparameterizable convolutional token mixing, integrated into user device operating systems, for on-device inference, capable of detecting and blocking harmful content without requiring internet or cloud connections, using lightweight machine learning models trained on diverse multimedia datasets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If cloud-based content filtering systems are used, then content detection capability is improved, but privacy risks and operational dependency increase
Solution Approach 1:
The patent extracts the content filtering functionality from cloud-based systems and embeds it directly into the user device. The machine learning model is integrated into the device's operating system, allowing content detection to occur locally on the device rather than requiring cloud connection, thereby eliminating privacy risks associated with cloud processing while maintaining detection accuracy
Solution Approach 2:
The patent introduces an intermediate layer - the on-device machine learning model - that mediates between the content to be filtered and the user device. This intermediary enables accurate content detection to occur locally without requiring direct cloud connection, resolving the contradiction between detection accuracy and privacy protection
2Reliability
If comprehensive content filtering is implemented, then content safety is improved, but device resource consumption increases
Solution Approach 1:
The patent implements partial action by activating the content filtering mechanism selectively rather than continuously. The system monitors content and activates the machine learning model only when potential harmful content is detected or when appropriate based on configuration, thereby maintaining content safety while reducing unnecessary battery consumption during normal device operation
Solution Approach 2:
The patent changes the operational parameters of the machine learning model to optimize for mobile device constraints. The model is trained and configured to operate efficiently with limited computational resources, adjusting processing intensity and activation thresholds to balance content safety requirements with battery conservation
3Measurement precision
If advanced AI models are deployed, then detection accuracy is improved, but device complexity increases
Solution Approach 1:
The patent segments the AI system into modular components: the machine learning model is separated from the operating system core but integrated through defined interfaces. This segmentation allows advanced detection capabilities to be added without fundamentally complicating the overall system architecture, as the AI component functions as a distinct, manageable module within the device
Data Source
AI summary
An embedded artificial intelligence system for preventing exposure to, and the creation of, inappropriate content by automatically detecting and blocking harmful content, including self-generated content. The system includes a machine learning (ML) model, built on efficient vision transformer (EVT) architecture, wherein the EVT architecture is operable to balance high performance with low computational overhead, such that the system is operable to be embedded in the operating system of a device. The ML model is trained on a dataset including images, text descriptions, videos, audio and/or other multimedia content, wherein the dataset content includes neutral images and/or inappropriate images, including not safe for work (NSFW) images and illegal child sexual abuse material (CSAM). The system is operable to detect and block content being shown on a device, camera-captured content taken on a device, and/or camera-captured content broadcasted in real-time/livestreamed from a device.