Interpretable Rotation-Equivariant Multiary-Valued Network for Attribute Obfuscation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Neural networks can leak private input information through intermediate-layer features.
• Attackers may invert these features to reconstruct sensitive data.
• Existing methods for preventing leakage can be computationally expensive or complex.

Purpose of the Study:

• To propose a generic method for preventing information leakage in neural networks.
• To introduce a novel multi-ary valued rotation-equivariant neural network (RENN).
• To ensure that the proposed method maintains the utility of neural network features for downstream tasks.

Main Methods:

• Converting real-valued features into multi-ary features.
• Hiding input information within the phase of multi-ary features.
• Applying rotation transformations (with private keys) for attribute obfuscation.
• Designing RENN to be rotation-equivariant by revising classic neural network operations.

Main Results:

• RENN effectively prevents information leakage even when network parameters and features are compromised.
• The encryption process preserves spatial correlations, allowing seamless integration with convolution operations.
• Experimental results show significant improvement in preventing information leakage with only mild degradation in classification accuracy.
• The computational cost of RENN is substantially lower than homomorphic encryption.

Conclusions:

• RENN offers a robust and efficient solution for preventing information leakage in neural networks.
• The rotation-equivariance property is crucial for maintaining feature utility post-encryption.
• RENN presents a practical approach for enhancing data privacy in machine learning models without compromising performance significantly.