Engravings, Secrets, and Interpretability of Neural Networks

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Neural networks (NNs) are increasingly vital in society.
• Ensuring the integrity of NN training and usage is critical.
• The problem of undetectable information embedding in NNs requires systematic investigation.

Purpose of the Study:

• To define security criteria for secret information engravings in NNs.
• To develop machine learning methods for creating engraved NNs.
• To establish a threat model for evaluating engraving security using interpretability methods.

Main Methods:

• Machine learning training algorithms for constructing engraved NNs.
• Development of distinguisher algorithms based on a defined threat model.
• Evaluation of engraving resilience against state-of-the-art interpretability methods on image classification datasets.

Main Results:

• Proposed definitions for the security of secret engravings in NNs.
• Demonstrated machine learning constructions for engraved networks.
• Identified specific NN engravings vulnerable to proposed distinguishers.
• Showcased resilience of systematic engravings against distinguishing attacks on benchmark datasets.

Conclusions:

• The developed threat model provides a benchmark for NN interpretability methods.
• Systematic engraving methods offer resilience against detection and attacks.
• This work advances the security and trustworthiness of neural networks.