AI Model Capture Prevention via Blocker Module
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI systems are vulnerable to model stealing attacks where adversaries can capture and replicate AI models by sending iterative queries, leading to potential business disadvantages, loss of confidential information, and intellectual property theft.
Innovation Solution
An AI system comprising an input interface, a blocker module, an AI module, and a notification module that detects and manipulates attack input data to prevent model capture, using unsupervised learning to train the blocker module to differentiate between attack and valid input data, and actively manipulating output to mislead attackers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the AI system uses a trained model to process input data, then the system can generate accurate output based on learned intelligence, but the model becomes vulnerable to capture by adversaries who can extract internal information through iterative queries
Solution Approach 1:
A query engine is introduced as an intermediary component between the adversary and the trained model. The query engine modifies output vectors by inserting queries into the model's decision logic, acting as a mediator that preserves accurate classification for legitimate users while preventing adversaries from extracting internal model information through iterative queries
Solution Approach 2:
The system converts the harmful effect of model transparency into a benefit by using the query engine to disguise the trained configuration of the model logic. The same mechanism that enables accurate classification is used to obfuscate internal model structures from adversaries, turning potential vulnerability into protection
2Measurement precision
If the system processes large amounts of training data to develop accurate AI models, then the model performance improves, but the time and resources invested in development are lost if the model is captured and replicated by adversaries
Solution Approach 1:
The query engine performs preliminary anti-action by preemptively blocking adversaries from capturing the model. Before adversaries can extract internal information through iterative queries, the query engine intercepts and modifies the interaction, preventing the theft of development time and resources invested in training the model
3Ease of operation
If the AI system maintains transparency in model logic for debugging and improvement, then the system can be easily maintained and refined, but adversaries can more easily extract and replicate the model logic
Solution Approach 1:
The system applies different qualities to different parts of the model interaction. The query engine maintains full transparency and accessibility of model logic where needed for debugging and improvement, while simultaneously applying obfuscation and modification where adversaries attempt to extract information, creating local variations in transparency based on the nature of the query
Data Source
AI summary
An AI system includes an input interface configured to receive input data from at least one user, and a blocker module configured (i) to detect the received input data is an attack input data, (ii) to manipulate output data of the blocker module, and (iii) to send the output data from the blocker module to the at least one user. The AI system further includes an AI module configured to process the input data received by the blocker module and to generate output data corresponding to the input data. The AI system also includes a blocker notification module configured to transmit a notification to the owner of the AI system on detecting an attack input data, and an output interface configured to send the generated output data to the at least one user.