ShrewdAttack: Low Cost High Accuracy Model Extraction.

Area of Science:

• Cloud Computing Security
• Machine Learning Security
• Artificial Intelligence

Background:

• Machine Learning as a Service (MLaaS) enables enterprises to utilize pre-trained models, streamlining business operations.
• Model extraction attacks pose a significant threat by allowing attackers to steal MLaaS model functionality.

Purpose of the Study:

• To develop a novel, cost-effective, and accurate model extraction method for MLaaS.
• To address the security vulnerabilities introduced by readily available MLaaS models.

Main Methods:

• Utilized pre-trained models and task-relevant data to minimize query data size.
• Implemented instance selection to reduce the number of query samples.
• Categorized query data into low-confidence and high-confidence sets to optimize budget and accuracy.

Main Results:

• Achieved high substitution accuracy (96.10% and 95.24%) on Microsoft Azure models.
• Required querying only a small fraction of the training data (7.32% and 5.30%).
• Demonstrated a low-cost and high-accuracy model extraction technique.

Conclusions:

• The proposed method presents a significant security challenge for cloud-based machine learning models.
• Highlights the urgent need for advanced mitigation strategies to protect MLaaS platforms.
• Future research may involve generative adversarial networks and model inversion for enhanced attack data generation.