Model-Inversion-Resistant Physical Unclonable Neural Network Using Vertical NAND Flash Memory

Area of Science:

• Computer Science
• Electrical Engineering
• Cryptography

Background:

• Growing use of neural networks in sensitive applications requires enhanced data and model protection.
• Existing architectures are vulnerable to model-inversion and cloning attacks, compromising privacy and intellectual property.

Purpose of the Study:

• To present a novel model-inversion-resistant physical unclonable neural network (PUNN) implemented on commercial vertical NAND (V-NAND) flash memory.
• To demonstrate hardware-rooted security against model-cloning and model-inversion attacks while maintaining high accuracy.

Main Methods:

• Implemented a PUNN using weak gate-induced drain-leakage erase on V-NAND flash memory to create unique, unreproducible device-level conductance patterns.
• Utilized the forward-forward (FF) algorithm for training, compatible with V-NAND's common-source-line structure, eliminating backward propagation.
• Evaluated model non-clonability by transferring trained weights between chips and assessed privacy preservation on the MIT-BIH electrocardiogram dataset.

Main Results:

• The V-NAND FF-PUNN demonstrated hardware-rooted resistance to model-cloning, with inference accuracy collapsing upon weight transfer to different chips due to inherent randomness.
• The system achieved high accuracy in classifying electrocardiogram data while completely preventing data reconstruction via model-inversion attacks.
• The PUNN maintained competitive classification accuracy under forward-only learning.

Conclusions:

• Established a scalable framework for secure, energy-efficient, and privacy-preserving neural computing directly on commercial flash memory.
• The V-NAND FF-PUNN offers a practical solution for protecting sensitive data and model integrity in real-world applications.
• This approach provides intrinsic non-clonability and robust defense against sophisticated cyber threats.