AI Module Query Blocking Against Model Capture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Adversaries can capture AI models by sending iterative queries, creating a digital twin or replica, which leads to business disadvantages, loss of confidential information, and intellectual property theft.
Innovation Solution
Implement a blocker module in the AI system that detects invalid input data, uses class swapping techniques to generate misleading outputs, and notifies the owner of potential attacks, thereby preventing model extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the AI system uses trained models to process data and generate output, then the system provides accurate classification and decision-making, but the models become vulnerable to capture and replication by adversaries
Solution Approach 1:
The patent introduces a query engine as an intermediary component between the adversary and the trained model. This query engine modifies the output vector by inserting queries into the model's internal functions, acting as a mediator that prevents direct access to the model's internal logic while still providing classification outputs. This intermediary layer blocks the adversary's ability to capture the model by intercepting and modifying the interaction between external queries and the protected model.
Solution Approach 2:
The patent changes the parameters of the output vector by modifying its values through the query engine. The query engine inserts queries that alter the output vector's content, transforming it from a direct reflection of the model's internal state to a modified output that maintains classification accuracy but conceals the model's true configuration. This parameter modification prevents adversaries from extracting meaningful information about the model's internals.
2Reliability
If the system processes large amounts of training data to develop robust models, then the model performance improves, but the development time and resources increase
Solution Approach 1:
The patent applies preliminary action by pre-training the model with large amounts of data before deployment. The model undergoes extensive training in advance to achieve robust performance, and then this pre-trained model is protected by the query engine. This allows the system to benefit from thorough training without exposing the time-consuming development process to adversaries, as the model is already trained and protected before any external interaction occurs.
3Object-affected harmful factors
If the query engine modifies the output vector to disguise the trained configuration, then the model protection improves, but the system complexity increases
Solution Approach 1:
The query engine serves multiple functions: it modifies the output vector to protect the model, maintains classification accuracy, and prevents model capture. By consolidating these multiple protective and functional responsibilities into a single query engine component, the patent reduces overall system complexity compared to having separate mechanisms for each function. The query engine's multi-functionality allows it to handle model protection, output modification, and accuracy maintenance within one integrated module.
Data Source
AI summary
An AI system and a method to prevent capture of an AI module in an AI system are disclosed. The AI system includes at least: an input interface to receive input data from at least one user; a blocker module to detect validity to the received input data and to generate output data corresponding to invalid data being received as input data using a class swapping technique; a data set module to store valid input data sets; an AI module to process the input data and generate output data corresponding to the input data; a blocker notification module to transmit a notification to the owner of the AI system on detecting an invalid input data and an output interface to send the generated output data to the at least one user.