AI Module Blocker Using Correlation Detection Against Model Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI systems are vulnerable to model extraction attacks, where adversaries iteratively query the system to extract internal information, leading to potential business disadvantages, loss of confidential information, and intellectual property theft.
Innovation Solution
Implement a blocker module in the AI system that calculates a correlation factor based on non-linear mutual information to detect and modify outputs, preventing unauthorized access and notifying the owner of potential attacks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the AI system processes and outputs results based on trained models, then the system provides useful service and generates value, but the system becomes vulnerable to model extraction attacks where adversaries can iteratively query the system to steal internal model information
Solution Approach 1:
The patent applies preliminary action by pre-training a detection model on normal query patterns before deployment. This detection model is then used to analyze incoming queries and predict potential attacks in advance, allowing the system to take preventive measures before model extraction occurs. The blocklist is also pre-computed based on detection model predictions, enabling proactive defense rather than reactive response.
Solution Approach 2:
The patent introduces a detection model as an intermediary component between the AI system and user queries. This detection model acts as a mediator that analyzes queries, predicts attack likelihood, and determines whether to block requests before they reach the core AI model. The blocklist serves as another intermediary layer that filters out malicious queries, protecting the underlying trained models from direct exposure to adversarial inputs.
2Reliability
If the system blocks potential attack vectors, then model security is improved, but legitimate queries may be incorrectly blocked reducing system availability
Solution Approach 1:
The patent applies parameter changes by adjusting the threshold parameter in the detection model to balance security and availability. By tuning this threshold, the system can control the sensitivity of attack detection, allowing legitimate variations in query patterns while blocking actual attacks. The block probability parameter is dynamically adjusted based on detection confidence, enabling flexible control over the trade-off between security and availability.
Solution Approach 2:
The patent implements feedback mechanisms where the detection model continuously learns from blocked and non-blocked queries. The system monitors detection accuracy and adjusts the blocklist and detection thresholds based on feedback from actual attack patterns and legitimate user behavior. This feedback loop enables the system to improve its discrimination between malicious and benign queries over time, reducing false positives while maintaining security.
Data Source
AI summary
A method to prevent exploitation of an AI module in an AI system is disclosed. The AI system includes an input interface, an output interface, an AI module, an information gain module, a blocker module, and at least a blocker notification module. An important non-limiting feature of the AI system is the functionality of the blocker module. The blocker module is configured to restrict at least one user by comparing a computed correlation factor amongst a plurality of inputs received from one or more users with a pre-determined range of correlation factors. The blocker module then modifies the output of the AI module in dependance of the comparison to restrict at least one user.