Research statement and record

Research

I work where computer vision, robotics and cyber-physical systems security meet.

In my last project I used a hobby laser to blind nine commercial smart cameras — Google Nest, Ring, Arlo and others — from as far as 120 metres away. Every device failed, and none of them told its owner. I then wrote two training-free algorithms that read the physics of the injected light instead of retraining a neural network. They catch the attack 98.44% and 99.22% of the time.

I filed the defences as provisional patents, disclosed the vulnerabilities to Google, Amazon and Arlo, and declined the bug bounties so the work could stay publishable.

Lab
Cyber-Physical Security Lab
Advised by
Dr. Yazhou Tu
Department
Computer Science and Software Engineering
School
Samuel Ginn College of Engineering, Auburn University
Since
Aug 2024

Publications

4 entries

Peer-reviewed and in-submission work

Each entry lists the authors in the order the venue records them, with my name in bold. Where a venue records equal first authorship, the line under the entry says so.

  • In submission

    2026

    Seeing Isn't Securing: Mitigating Vision Perception Threats in Smart Home Security Systems by Harnessing Adversarial Optical Physics

    Eftakhar Ahmed Arnob, Yazhou Tu, Xingli Zhang, Kevin Butler, Xiali Hei

    I tested nine commercial security cameras and doorbells from six manufacturers under laser light injection. Every one failed: person, face, package and even basic motion detection went quiet, while the app still reported nothing wrong. The paper then builds two defences that read the physics of the injected light instead of retraining the AI, and tests them across 970 scenario videos.

    CPS security · computer vision · optical injection

  • Accepted

    2026

    Plant-Inspired Elastic–Hydraulic Tactile Sensing Enables Quantitative Stiffness Estimation in Soft Robots

    Tofayel Ahammad Ovee, Eftakhar Ahmed Arnob, Jean-François Louf

    Tofayel Ahammad Ovee and Eftakhar Ahmed Arnob share equal first authorship.

    Soft Robotics (SAGE)

    A soft robotic finger that senses the way a plant does — fluid moving through an elastic, porous body. When the finger presses something, the pressure in an embedded liquid channel shifts by an amount that maps onto how stiff that thing is, so the robot measures stiffness by touch across more than two orders of magnitude. We 3D-print it for under $50. It can tell you whether a piece of fruit is ripe.

    soft robotics · tactile sensing · control

  • Under review

    2025

    Efficiency-Driven Shape Completion in Sparse Visual Domains

    Tofayel Ahammad Ovee, Eftakhar Ahmed Arnob, Jean-François Louf

    Tofayel Ahammad Ovee and Eftakhar Ahmed Arnob share equal first authorship.

    Preprint on SSRN · under review

    Give it a shape a sensor only partly saw and it predicts the rest, fast enough to sit inside a control loop. We designed the whole thing around that speed constraint rather than around accuracy on its own.

    machine learning · shape completion

  • Published

    2025

    ED² dataset

    Tofayel Ahammad Ovee, Eftakhar Ahmed Arnob, Jean-François Louf

    IEEE DataPort · CC BY 4.0

    The open dataset behind the Efficiency-Driven Encoder–Decoder work: hand-drawn sketches and occluded objects for shape completion. The model reaches 80–90% structural similarity even when three-quarters of the input is missing, and it uses a fraction of the compute a GAN or U-Net variant needs.

    dataset · open science

Patents

2 entries

Filed, not granted

Provisional filings. A provisional application records a priority date and nothing more; the line under each entry names exactly what I filed and where.

  • Filed

    Filed 29 Jan 2026

    Methods for Detecting Optical Injection Attacks on Cameras by Harnessing Adversarial Optical Physics

    Yazhou Tu, Eftakhar Ahmed Arnob, Xiali Hei, et al.

    USPTO provisional · App. 63/970,607 · Auburn IPX 2026-025

    The defence as I filed it: split the camera's field of view into a grid, compare each cell's colour channels against a baseline, and flag the signature a laser leaves and sunlight does not — one channel spiking while the camera's own auto-exposure pushes the other two down. It uses no model and no training data, so there is nothing to retrain.

    patent · detection

  • Filed

    2026

    Lens Shield — Out-of-View Laser / Optical Injection Shield

    Eftakhar Ahmed Arnob, Yazhou Tu, Xiali Hei, et al.

    Auburn IPX 2026-070

    A hardware answer to the harder version of the attack: a laser aimed from outside the camera's field of view, at an angle where the source never appears in the footage.

    patent · hardware

Artifacts

2 entries

Artifacts and datasets

Code and data I deposited. Each line links to the copy of record.

Talks

2 entries

Conference talks

Talks on work I co-authored. Each entry names the author who stood up and gave it, which is not always me.

  • Co-authored talk

    18–20 Dec 2023

    Plant-Inspired Mechanosensitive Soft Robots

    CompFlu 2023 · 17th edition

    IIT Madras, Chennai, India · Responsive Softmatter, Hall 2

    Presented by T. A. Ovee

  • Co-authored talk

    4 Mar 2024

    Plant-Inspired Mechano-Sensing Soft Robot

    APS March Meeting 2024 · DSOFT / DBIO

    Session D38, Robophysics I · Minneapolis, USA

    Presented by T. A. Ovee

Press

1 entry

In the press

One article, on the camera work. I quote the passages below exactly as the article prints them.

Cybersecurity nowadays extends farther than computers, websites or servers. We must also consider cyber-physical systems. If you look at drones, robotic dogs, smart homes, and autonomous vehicles, these systems are getting more intelligent to perceive the environment and handle things automatically.

Also from the article

The cameras also don't notify homeowners that their view is being blocked by the laser.

Assistant professor in CSSE exposes vulnerability in AI-powered security cameras, offers new defenses in development

Auburn Engineering · 10 October 2025

Written by Joe McAdory

Profiles

Where the record lives

You don't have to take any of this on trust. The registries themselves hold these entries.

The same record on one page: curriculum vitae (PDF).