Physical unclonable in-memory computing for simultaneous protecting private data and deep learning models.

Area of Science:

• Computer Engineering
• Hardware Security
• Artificial Intelligence

Background:

• Compute-in-memory (CIM) using resistive random-access memory (ReRAM) accelerates neural networks on edge devices, improving energy efficiency.
• The nonvolatile nature of ReRAM poses security risks, allowing potential extraction of sensitive neural network weights during computation.

Purpose of the Study:

• To propose a robust data protection scheme, RePACK, for secure neural network computation in ReRAM-based CIM systems.
• To safeguard neural network input data, stored weights, and structural information against unauthorized extraction.

Main Methods:

• Developed RePACK, a threefold data protection scheme integrating a bipartite-sort coding strategy.
• Incorporated a fully on-chip physical unclonable function (PUF) for enhanced security.
• Implemented and evaluated the RePACK system on a 40nm ReRAM CIM chip.

Main Results:

• Successfully protected neural network input, weight, and structural information.
• Demonstrated a significant increase in enumeration complexity to 5.77 × 10^75 for a 128-column CIM core.
• Validated the RePACK system's effectiveness on a physical hardware prototype.

Conclusions:

• RePACK provides a viable solution for enhancing the security of ReRAM-based CIM systems for edge neural networks.
• This work contributes to developing safe, robust, and efficient edge AI accelerators, potentially supporting applications like federated learning.