Off-campus SJSU users: To download campus access theses, please use the following link to log into our proxy server with your SJSU library user name and PIN.

Publication Date

Summer 2026

Degree Type

Thesis - Campus Access Only

Degree Name

Master of Science (MS)

Department

Electrical Engineering

Advisor

Chang “Charles” Choo; Mai-Khanh (Mike) Nguyen; Shrikant Jadhav

Abstract

This thesis designs and validates an end-to-end face verification pipeline that runs entirely on the MAX78000 TinyML microcontroller, stores no recoverable biometric template, and derives a deterministic 128-bit cryptographic key directly from the user's face. An 88×88 grayscale image is passed through a 76,688-parameter INT8 CNN that regresses eight facial landmarks, which are reduced to a 21-dimensional geometric feature vector and encoded as a 10,000-bit binary hypervector via hyperdimensional computing. A cancelable transform—a one-time-pad XOR with a hardware-random key hypervector followed by a Fisher–Yates bit permutation—wraps the hypervector so that a compromised template can be revoked and re-issued from the same face. A tiled BCH fuzzy extractor derives a stable 128 bit key from the noisy cancelable template, tolerating up to approximately 30% bit-flip noise; the key is never stored and is regenerated on every verification from stored public helper data. Validation over 1,000 randomized trials confirms that the cancelable scheme attains its information-theoretic bounds for irreversibility and unlinkability, that the fuzzy extractor achieves 100% reliability and 100% selectivity, and that the C firmware is bit-exact with the Python reference. The firmware occupies 192 KB of Flash and 33 KB of SRAM, with verification completing in 25–30 milliseconds at approximately one millijoule per match.

Download

Share

COinS