● Sylhet → Auburn, Alabama
Eftakhar Ahmed Arnob
I work on machines that sense the world — and on how they get fooled.
I started with a robotic arm and a car I could drive from the other side of the planet. Then a network that finishes drawings people only half-drew. I’m two years into a PhD at Auburn now, taking apart what security cameras think they can see.
Scroll to travel · drag the portrait
Building now · Auburn University
Attacking a self-driving car, on purpose.
This is the project I’m working on: a mixed-reality rig for testing how well self-driving systems and AI agents hold up when somebody is deliberately trying to break them. A real vehicle, real sensors, and adversaries we composite into what the car believes it’s looking at.
The idea is to find the failure in a lab, on a schedule, instead of finding it on a road.
Mid-test
Everything holds, until it doesn’t.
The car reads the road, the trees, the sign. Then we put something in front of it that was designed to be misread. Boxes drift off their objects. Labels change. Things it was tracking a moment ago stop existing, and a couple appear that were never there.
The trick isn’t the interesting part. What I want is the measurement: how much light, how much distance, how much perturbation — and which part of the stack gives way first.
Why it matters to me
Nobody fixes a failure they’ve never seen.
Perception stacks get tested against weather, traffic and awkward edge cases. They rarely get tested against somebody who is actually trying. I want that kind of testing to be routine — repeatable, instrumented, and dull enough to run every day.
2019 — 2023 · North East University Bangladesh
It started on a bench.
I meant the arm to be glove-controlled. The parts budget said otherwise, so it got a control panel instead. I built a car I could drive over the open internet from anywhere, and a pump that watched a water tank so nobody else had to.
In between I did three years of ICPC, won my university’s contest, and ended up running the robotics workshops I had signed up to attend.
Undergraduate thesis · the ED² dataset
Give it half a shape.
I trained an encoder–decoder to finish shapes somebody only partly drew. It holds 80–90% structural similarity with three quarters of the input missing, and it runs on a fraction of what a GAN needs. The dataset is public.
Draw on the panel. It’ll finish what you start.
- Circle
- Rectangle
- Triangle
- Star
- Heart
The five classes in the dataset
2024 — now · Auburn University
Nine cameras, one laser.
Point a cheap laser at a smart doorbell and its AI stops seeing — person, face, package, even plain motion. I tested nine devices from six manufacturers. All 28 of 28 detections failed under static injection, from up to 120 m away, and not one camera told its owner anything had happened.
Then I wrote the defence: two training-free checks that read the physics of the injected light instead of retraining a model. The readout below is running on this frame.
The back catalogue
Everything I’ve built.
Robotic arms, an institute management system, a social network I wrote from scratch in PHP, a forum for my university’s programmers, a 2D game in Java Swing, control rigs for soft robots. Some of it is research. Most of it was curiosity.
- 04
- Research systems
- 03
- Hardware builds
- 05
- Software projects
Contests
Where I competed.
Three years of ICPC regionals in Dhaka, a Code Jam qualification, a hackathon podium and a university championship. I’m not the strongest competitive programmer you’ll meet, but the habit of reading a problem properly before touching the keyboard started here.
- National 90th, 103rd, 141stICPC Asia Dhaka Regional, onsite2020 · 2021 · 2022
- World 1677th · National 13thGoogle Code Jam, qualification round2021
- 2nd runner-upSJ Innovation Hackathon, team Gladiators2021
- ChampionNEUB Eid Mubarak Programming Contest2019
- 1st positionNational Talent Hunt Competition2014
Paper trail
The certificates.
The actual scans, in case you would rather see them than take my word for it. Move your cursor across them.
- Honourable mentionICPC Asia Dhaka Regional Contest 20212021
- Honourable mention2020 ICPC Asia Dhaka Regional Contest28–29 May 2021
- Honourable mentionICPC Asia Dhaka Regional Preliminary7 April 2021
- Honourable mentionICPC Asia Dhaka Regional Preliminary16 July – 28 September 2019
- Qualification round · 90 pointsGoogle Code Jam 20212021
- 2nd runner-upSJ Innovation Hackathon 20217 October 2021
That’s the tour
Say hello.
I’m looking for collaborators, reviewers, and people working on sensing systems that might break in interesting ways. If any of this overlaps with what you do, I’d like to hear from you.
Computer Science and Software Engineering · Auburn University