A hardware root of trust for ESP32 chips with no dedicated TPM and secure elements, built from their SRAM fingerprint alone.
100% key reproducibility2 attacks, then hardened
open log ↵
Cheap microcontrollers like the ESP32 ship with no TPM or secure element. Their SRAM still carries a fingerprint: the value each cell settles into at power-up, set by tiny manufacturing differences. I built a hardware root of trust on that fingerprint, then threat-modeled and attacked it.
data512-bit responses over 10 cold boots on two boards: about 6% noise per boot, and 19–21% of cells metastable.
resultAn ECC threshold of t ≈ 110 gave 100% key reproducibility and 55.47% inter-device Hamming distance between derived keys.
dataThe metastable cells barely overlap between boards: 28 shared positions, a Jaccard distance of 0.841.
vulnThreat-modeled the prototype and demonstrated two attacks: cloning the identity by dumping the stored golden baseline from flash, and replaying its static authentication hash.
fixHardened it with a code-offset secure sketch (no key at rest) and HMAC-SHA256 nonce challenge–response.
limitQuantified helper-data leakage: about 990 bits at t = 110, n = 512, so a 512-bit array keeps no secret after error correction. A 256-bit key needs at least 2 KB of SRAM.
nextMasking unstable cells at enrollment and concatenated codes, measured on real ESP32 data.
Manuscript in preparation with Prof. Alon Hillel-Tuch [M1].
[2] llm4security · RTL
Researcher · Prof. Ramesh Karri, Weihua Xiao · 2026–now
Do multi-agent LLM pipelines find hardware security bugs better, or just spend more compute?
+25 pts over one agent84 benchmark cases
open log ↵
A controlled study: the model, task, scorer, and a 24-call budget stay fixed while only the orchestration changes, so any gain comes from the agent design and its tools.
setupFour configurations: tool-free reading, one agent with EDA tools, a supervisor–analyst–verifier pipeline, and that pipeline with a blind critic.
result80-run pilot: the pipeline beat a single tool-using agent by 25 points (90% vs. 65%; 95% CI +10 to +40).
negIt didn’t clearly beat careful reading without tools, and model-written checkers raised false alarms on 58% of safe-design verdicts.
bench84 oracle-labeled RTL cases, hash-frozen; 64% from MITRE CWE examples, Trust-Hub Trojans, and Hack@DAC 2021 SoC bugs.
stackA 4-bit 27B model on vLLM/ROCm across two AMD Radeon AI PRO R9700 GPUs. Icarus Verilog, Verilator, and Yosys/ABC run as MCP servers in network-isolated containers, and every tool call lands in a hashed evidence log.
nextA pre-registered five-hypothesis study: 2,268 runs per model, Holm-corrected tests, replicated on a second model family.
Manuscript in preparation with Weihua Xiao and Prof. Ramesh Karri [M2].
[M1]Hardware-Based Authentication for Edge Devices using SRAM Physical Unclonable Functions.V. Valle, A. Hillel-Tuch.in prep
[M2]LLM4Security: A Controlled Evaluation of Multi-Agent LLM Orchestration with EDA-Grounded Evidence for Hardware-Security Verification.V. Valle, W. Xiao, R. Karri.in prep
[P1]System for Assisting a Visually Impaired User.B. A. Baje, V. M. Q. Valle. Utility models PH 2-2020-050872 and JP 3239665; WO 2021/230759 A1.registered
pre-university papers (3)
[T1]Wearable Real-Time Respiratory Disease Detection System with Convolutional Neural Networks Using Sound Data.S. S. D. Capadngan, V. M. Q. Valle. Research thesis, Philippine Science High School, 2022.
[R1]Development of a Visually Impaired-specific System for Indoor and Outdoor Navigation.V. M. Q. Valle, R. M. G. Alidon, S. S. D. Capadngan. Singapore International Science Challenge, 2021.
[R2]VISION: Visually Impaired-specific System for Indoor and Outdoor Navigation.V. M. Valle. IDEX 2020.
A privacy check in two parts. The first is a browser fingerprint demo: it shows what your browser gives away about you on its own. The second is an OSINT exposure check: it searches popular social media sites for a username and shows how exposed that handle is.
Lightweight Linux distros that ricing and the internet more accessible during COVID.
#linux#ricing
open ↵
Free, lightweight Linux distributions for old, low-end computers. During COVID-19 they let about 5,000 students attend online classes on decades-old machines.
The lab’s cluster, self-hosted LLMs, and zero-trust stack.
#proxmox#zero-trust
open ↵
As development lead of NYU’s OSIRIS Lab, I built a 3-node Proxmox cluster with Ceph and ZFS for research and CTF infrastructure, and a self-hosted LLM service (vLLM, Ollama, Open WebUI) on two RTX 3090s that keeps inference on-premises.
I secured it with Twingate zero-trust access, UniFi IPS/IDS, a Wazuh SIEM, and Authentik identity management, and remediated XSS (CVE-2026-24778) and SQLi (CVE-2026-26980) exposure in the lab’s Ghost CMS.
csaw-ctf
2024
Reversing and forensics challenges built to resist LLM-assisted solving.
#ctf#reversing
open ↵
Reverse-engineering and forensics challenges for the CSAW ’24 CTF Qualifiers, written to resist LLM-assisted solving. I also playtested challenges for CSAW ’23 and ’24.
As CSAW (Cybersecurity Awareness Week) lead at the NYU Center for Cybersecurity, I also support CI/CD, outreach, and sponsorships.
python-dlib
2024–2025
Co-maintained the Arch package for dlib’s Python bindings.
#arch-linux#packaging
open ↵
Co-maintained python-dlib on the Arch User Repository: the Python bindings for dlib, a C++ toolkit of machine-learning algorithms and tools.
hardware
vision
2019–2022
A wearable navigation aid for the visually impaired with all computer vision and AI inferencing done locally on-device
#edge-ai#assistive-tech#pcb
open ↵
A wearable navigation aid for the visually impaired that processes all computer vision on the device, so no image leaves the wearer. I founded and led the project.
Custom PCB with a Raspberry Pi CM3+, an Intel Neural Compute Stick 2, a PiCam V2, GPS, an MPU6050, and a voice assistant. 157.5 ms worst-case latency in dynamic scenes, 13–14× faster than a stock Raspberry Pi 3B+.
Registered as a utility model in the Philippines and Japan, with PCT application WO 2021/230759 A1. Presented at IDEX 2020 and the Singapore International Science Challenge 2021.
A clip-on wearable that classifies respiratory diseases with an on-device convolutional neural network.
#edge-ai#audio
open ↵
A clip-on wearable that classifies lung sounds with a transfer-learned Inception V3 CNN on spectrograms, for early detection of respiratory diseases. Compiled with OpenVINO for an edge accelerator: 57 ms per inference and 89% model accuracy.
I designed the electret microphone amplifier and the custom PCB. It became my research thesis with S. S. D. Capadngan [T1].
reefpatrol
2020
A LoRa buoy that streams coral-reef water quality to an open-data app.
#lora#sensors
open ↵
A LoRa-linked buoy that streams pH, turbidity, temperature, and dissolved oxygen from an Arduino Nano on a custom PCB to an open-data Android app, for coral-reef monitoring.
I led the team and presented it at the 8th Asian Youth Innovation Awards.
hikaway
2020
An Arduino particulate-matter monitor for asthma pre-detection.
#arduino#health
open ↵
An Arduino-based particulate-matter monitor for asthma pre-detection, built to help prevent asthma attacks. I led the hardware; the sensors reached 99% accuracy across 6 trials.
robomaster
2024
A CAN interface for NYU RoboMaster’s controls team.
#pcb#embedded
open ↵
A CAN module that lets NYU RoboMaster Team Ultraviolet’s controls team interface with a PIC32CM microcontroller, designed in Altium Designer. Part of NYU’s Vertically Integrated Projects.
vaughn@vallev:~$ cat experience.log
experience
Sep 2026–nowGraduate TA, Application Security· NYU Tandon
For MS Cyber Fellows (online) and in-person MS students: secure C, web, database, cloud, and mobile security. Weekly office hours, grading, and the course’s GitHub submission pipelines.
Sep 2026–nowLead Researcher, LLM4Security· NYU
With Prof. Ramesh Karri and Weihua Xiao. A controlled evaluation of multi-agent LLM orchestration for hardware-security verification.
Summer 2025Undergraduate Summer Research Program· NYU Tandon
Jan 2025–Jan 2026Development Lead· OSIRIS Lab
Built and secured the lab’s cluster, self-hosted LLM service, and identity and monitoring stack; remediated XSS and SQLi exposure in the lab’s Ghost CMS.
Sep 2024–nowCSAW Lead· NYU Center for Cybersecurity
Cybersecurity Awareness Week. CTF challenge writing and playtesting, CI/CD, outreach, and sponsorships.
Sep 2024–nowSystems Administration Head, EG1004· NYU General Engineering
Primary incident-response contact for 35 departmental endpoints (Cortex XDR, InsightVM, BeyondTrust, BigFix, ESXi); coordinate Windows 11 patching with NYU IT.
Jun 2024–Sep 2025President· NYU Cybersecurity Club
Weekly CTF workshops and Hack Nights; demos of Android privilege escalation and Active Directory/Kerberos attacks; a bug-bounty and disclosure program with Bugcrowd.
Spring 2024Controls Team· NYU RoboMaster Team Ultraviolet
A CAN module for a PIC32CM microcontroller, designed in Altium.
Nov 2023–nowUndergraduate Researcher· OSIRIS Lab
With Prof. Alon Hillel-Tuch. A hardware root of trust from SRAM PUFs for ESP32 microcontrollers that have no TPM or secure element.
Sep 2023–nowUndergraduate TA· NYU General Engineering
Arduino and Raspberry Pi labs, digital logic, and CAD; mentored 7 accessibility-focused design teams.
2022–2023Peer Mentor· CAUSE Philippines
Guided Filipino students through U.S. and U.K. college applications; mentees were admitted to NYU and Amherst College.
2019–2022VP & Lead Hardware Engineer· EINSTEIN Club, PSHS
Led 6 humanitarian inventions and 4 innovation groups; schematics and PCBs in Autodesk EAGLE.
2018–nowFounder· Mirth OS and NixLabs
Open-source Linux distributions for low-end hardware.