Is Homomorphic Encryption-Based Deep Learning Secure Enough?

Area of Science:

• Computer Science
• Cryptography
• Machine Learning

Background:

• Increasing data collection in machine learning, especially deep learning, raises significant user privacy concerns.
• Homomorphic encryption offers a solution by enabling computations on encrypted data, with applications in finance and healthcare.
• The security of deep learning services utilizing homomorphic encryption remains an open question.

Purpose of the Study:

• To investigate the feasibility of user data privacy breaches in homomorphic encryption-based deep learning services.
• To propose and validate novel attack methods targeting security vulnerabilities in these systems.
• To assess the practical threat posed by these attacks in real-world scenarios.

Main Methods:

• Proposed three distinct attack methods: adversarial attack (communication link), reconstruction attack (input/output data), and membership inference attack (malicious insider).
• Designed and executed experiments to evaluate the effectiveness of these attacks.
• Simulated real-world exploit scenarios in financial and medical service contexts.

Main Results:

• Demonstrated that adversarial and reconstruction attacks pose a practical threat to homomorphic encryption-based deep learning models.
• Adversarial attacks significantly reduced average classification accuracy from 0.927 to 0.043.
• Reconstruction attacks achieved an average reclassification accuracy of 0.888.

Conclusions:

• Homomorphic encryption-based deep learning services are vulnerable to sophisticated privacy attacks.
• The proposed attacks highlight critical security gaps that need addressing for secure implementation.
• Further research is needed to develop robust defenses against these identified privacy threats.