A novel obfuscation method based on majority logic for preventing unauthorized access to binary deep neural networks.

Area of Science:

• Computer Science
• Electrical Engineering
• Cybersecurity

Background:

• Deep learning models, particularly deep neural networks (DNNs), are increasingly valuable assets requiring protection against unauthorized access.
• The rise of binary neural networks (BNNs) presents unique challenges and opportunities for hardware implementation and security.

Purpose of the Study:

• To propose a novel key-based algorithm-hardware co-design methodology for securing DNN models.
• To develop a protection strategy specifically tailored for BNNs while ensuring broad applicability to other neural network accelerators.

Main Methods:

• An innovative key-based algorithm-hardware co-design approach was developed to protect DNN models.
• The methodology was validated using post-layout simulations with Cadence Virtuoso on TSMC 40nm CMOS technology.
• Security was assessed against various attacks, including Boolean satisfiability, structural, reverse engineering, and side-channel attacks.

Main Results:

• The proposed method significantly reduces model accuracy with incorrect keys, effectively preventing unauthorized access.
• The approach demonstrated superior efficiency compared to existing solutions across different BNN architectures and datasets.
• The design achieved substantial reductions in area (43%), average power (79%), and weight modification energy (71%).

Conclusions:

• The key-based co-design methodology offers a robust and efficient solution for protecting DNN models, particularly BNNs.
• The approach enhances hardware security for neural network accelerators against diverse cyber threats.
• The validated design provides significant improvements in resource utilization and energy efficiency.