Machine Learning Resistant Amorphous Silicon Physically Unclonable Functions (PUFs)

TL;DR

This work addresses the vulnerability of many PUFs to ML modeling by introducing nonlinear, integrated a-Si:H photonic PUFs built from a ray-chaotic cavity with strong $ ext{Kerr}$ nonlinearity. The authors fabricate CMOS-compatible devices and generate a large challenge-response library using a spectrally encoded ultra-fast optical scheme, then evaluate multiple ML attacks on the Hadamard-transformed responses, including linear regression, kNN, decision-tree ensembles, and deep neural networks. A private-information metric $PV(p,q)$ based on a symmetric KL divergence is used to quantify residual security, with DNNs providing the best resistance yet leaving substantial information unexploited, especially as pulse energy increases and nonlinearity grows. They also present theoretical bounds for linear optical PUFs, showing that linear collapse limits the CRP space and shows the necessity of nonlinearity for robust, integrated PUFs. Overall, the study demonstrates a CMOS-friendly, nonlinear photonic PUF with meaningful ML resistance and practical implications for secure on-chip communications, while outlining paths for deeper integration and further security analysis.

Abstract

We investigate usage of nonlinear wave chaotic amorphous silicon (a-Si) cavities as physically unclonable functions (PUF). Machine learning attacks on integrated electronic PUFs have been demonstrated to be very effective at modeling PUF behavior. Such attacks on integrated a-Si photonic PUFs are investigated through application of algorithms including linear regression, k-nearest neighbor, decision tree ensembles (random forests and gradient boosted trees), and deep neural networks (DNNs). We found that DNNs performed the best among all the algorithms studied but still failed to completely break the a-Si PUF security which we quantify through a private information metric. Furthermore, machine learning resistance of a-Si PUFs were found to be directly related to the strength of their nonlinear response.

Figures (11)

• Figure 1: a) Scanning electron microscope (SEM) image of an amorphous silicon ray chaotic cavity used as a PUF. b) The device is probed by patterning a chirped pulse with the binary challenges and the output time series signal is measured in the Hadamard basis. c) Finite difference time domain (FDTD) simulation of a mode in the PUF device.
• Figure 2: Ultrafast optical setup for creating the CRP (challenge-response pair) library. Pulses from the MLL (mode locked laser) are stretched using dispersive fiber. Pulses are then patterned by an MZM (Mach-Zehnder modulator) modulated by a PPG (pulse pattern generator) triggered by the 8th harmonic of the MLL. The pulses are then compressed with equal but opposite amounts of dispersion. They are then fed through an EDFA (Erbium doped fiber amplifier), PC (polarization controller), PUF (physically unclonable function), EDFA again, and finally a programmable spectral filter.
• Figure 3: a) Average fractional Hamming distance (FHD) between PUF responses as well as linear regression guess as a function of the bit level (from the most significant bit to the least significant bit). Different colors correspond to different pulse energies (blue: 5pJ, green: 1.5 pJ, red: 0.5 pJ). Circles correspond to FHD between PUF responses probed twice for the same challenges and is a measure of the repeatability of the measurement. Triangles correspond to FHD between ground truth PUF response and the linear regression guess. LIN stands for linear model. b) Histogram of FHDs for the 5 pJ pulse energy for a 20kbit key length both for the PUF and the linear model prediction.
• Figure 4: Average fractional Hamming distance as a function of training data set size for bit levels 1-4. Training set size beyond a few thousand samples provides diminishing returns.
• Figure 5: Average FHD at 3 different power levels as a function of the bit level. Triangles correspond to kNN predictions and circles correspond to PUF ground truth (displayed for comparison). b) FHD histogram at each bit level for a key length of 20kbits.