Researchers Showcase Adversarial Patterns to Disrupt Surveillance Recognition

CoinWorld reported:

A cybersecurity professional from the United States showcased an "adversarial pattern" at the Def Con conference in Las Vegas. These computer-generated patterns can be printed on clothing, objects, or vehicle surfaces to interfere with the automatic recognition systems of certain surveillance cameras, making it difficult to identify people, vehicles, or facial information.

The project is called noRecognition. Developer Bill Swearingen told TechCrunch that he has conducted approximately 31 million experiments over the past year, aiming to generate patterns that can bypass common surveillance detection algorithms. According to him, these patterns do not prevent cameras from recording but disrupt the system's ability to recognize and alert on targets in the footage.

Def Con Completes First Public Test
Swearingen covered the latest pattern on a 2009 Toyota Yaris during the first public demonstration at Def Con, testing it against a Flock camera system. He stated that the demonstration proved the pattern's effectiveness in real-world environments, although the wheel area remained challenging to fully address. Donut Media, which assisted in filming, announced that the demonstration video will be released in the coming weeks.

He mentioned that these patterns can now be generated on demand and can be applied to clothing, hoodies, and potentially vehicle wraps in the future. The project team has also launched a crowdfunding campaign to sell early products featuring the related patterns.

The Goal is to Disrupt Automatic Detection, Not to Obscure the Footage
Swearingen explained that many current surveillance systems have target detection capabilities, which can be used to recognize license plates, track vehicles, or screen specific subjects through facial recognition. Unlike traditional obfuscation, this method does not make the target disappear from the video; instead, it prevents the algorithm from reliably determining what appears in the footage.

According to him, once the system fails to trigger automatic detection, the target will not be easily filtered out from vast amounts of video, requiring manual searching.

Model Generates New Patterns Through Trial and Error
Swearingen stated that the project initially began with a proof of concept using open-source video detection algorithms, gradually increasing computational power and evolving into a reinforcement learning model. This model continuously attempts different patterns; as long as a certain pattern is still recognized by the algorithm, it continues to iterate until it bypasses multiple detection models simultaneously.

He noted that the system has currently tested 11 open-source detection algorithms and has found combinations of patterns that can disrupt multiple algorithms at once, including those related to Flock's license plate recognition, Axon's law enforcement recording devices, and Clearview AI's software capabilities. According to him, the model now generates a batch of new patterns almost every minute, and the effectiveness continues to improve.

Swearingen stated that he has not yet disclosed the strongest version of the pattern to prevent related companies from quickly patching their recognition systems. However, he believes that this public demonstration has shown that using adversarial samples to evade algorithm detection is no longer confined to the laboratory stage in real public spaces.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.