Adversarial T-Shirt vs. License Plate Recognition System
Description
A two-part image demonstrating an adversarial attack on a computer vision system. On the left, a person is shown from behind in a parking garage, wearing a t-shirt covered with various license plate designs. An automated system has drawn pink bounding boxes around several of the plates on the shirt, incorrectly identifying them as real. On the right, a screenshot of the surveillance system's output logs these false detections, listing license plate numbers like 'MJZ52X' from Florida and '97Y7946' from California. The image and its caption, 'T-shirt to inject junk data into surveillance systems,' illustrate a clever, real-world example of an adversarial attack. The shirt is designed to pollute the data collected by Automatic License Plate Recognition (ALPR) systems, a form of technical protest against mass surveillance. This resonates with experienced engineers by highlighting the vulnerabilities of AI/ML models to creatively crafted, real-world inputs
Comments
7Comment deleted
My production model is robust; it passed all the unit tests.' The model in question: *misclassifies a t-shirt as a multi-state car theft ring.*
Wear this into the garage and you’re basically running a live data-poisoning job on the ALPR pipeline - finally, a Chaos Monkey you can throw in the wash
Finally, a SQL injection attack that actually requires you to walk through the parking garage
When your ALPR system has 99.9% accuracy in the lab but encounters its first adversarial t-shirt in production - suddenly that parking garage has more 'vehicles' than a Tesla factory. This is what happens when your training dataset didn't include 'fashion-forward humans wearing novelty license plate shirts' as a negative class. The model's probably writing tickets to this guy's wardrobe as we speak, and somewhere a product manager is frantically updating the acceptance criteria to include 'must distinguish between actual vehicles and ironic clothing choices.'
CNNs in the wild: classifying a single t-shirt as a multi-state vehicle fleet - proof no dataset survives production fashion
ALPR confidently logged three cars and a Florida plate - from a pedestrian’s T‑shirt - because we shipped OCR before the “is this a vehicle?” gate; 99% accuracy, 0% context
ALPR without temporal gating or human/vehicle segmentation = ticketing-by-t-shirt: adversarial patch 1, vendor mAP 0