Monolithic 3D Integration of Vertical Memory with Phototransistor for Near-Sensor Cryptography and Homomorphic Data...

Area of Science:

• Materials Science
• Computer Engineering
• Cryptography

Background:

• Retinomorphic systems offer efficient near-sensor processing by integrating sensing, memory, and computing.
• Biological vision lacks inherent security, posing challenges for artificial edge systems requiring data confidentiality.
• Next-generation hardware needs to merge bio-inspired efficiency with cryptographic functions for secure edge AI.

Purpose of the Study:

• To propose a compact, multifunctional monolithic 3D (M3D) architecture for secure in-memory processing of visual data.
• To integrate physical unclonable function (PUF) keys for enhanced cryptographic resilience.
• To achieve efficient and secure machine vision at the edge.

Main Methods:

• Integration of quantum dot-sensitized phototransistors with vertical resistive random-access memories (VRRAMs) for PUF key generation.
• Implementation of a multi-layer encryption scheme using diverse PUF keys.
• Development of M3D ternary content-addressable memory (TCAM) using wide-bandgap IGZO transistors.

Main Results:

• Achieved ≈50% inter-device variability for PUF keys, enhancing cryptographic security.
• Demonstrated 9.61× area efficiency and 6.25× energy-delay product improvement in M3D TCAM arrays.
• Enabled near-sensor hashing and in-memory computation on encrypted data with 94.1% similarity preservation.

Conclusions:

• The proposed M3D architecture effectively combines bio-inspired efficiency with cryptographic capabilities for secure edge AI.
• M3D systems support privacy-preserving machine vision by enabling direct computation on encrypted data.
• This platform holds significant potential for secure, efficient, and privacy-preserving machine vision applications at the edge.