New adversarial mask designs evade facial recognition systems
Fabric face masks covering the nose and mouth and printed with adversarial patterns evade facial recognition systems more than 96 percent of the time, find Israeli researchers.
Initially stumped by the mask-wearing prompted by the COVID-19 pandemic, facial recognition systems soon caught up (although with errors). Researchers at Israel’s Ben-Gurion University of the Negev and Tel Aviv University decided to take on the adversarial role to see if they could develop a specific pattern or mask that would work against current deep learning facial recognition models, reports Help Net Security.
Participants were asked to walk down a corridor while wearing various control masks such as a blue surgical mask and masks printed with realistic human features (of the wearer’s sex and opposite sex). A short video shows the process. In all scenarios, the participants’ faces were detected and recognized.
When wearing the adversarial pattern printed on both paper and fabric masks, the facial recognition algorithms detected a face, but could not make a match, meaning wearing one would not raise suspicion, reports Help Net Security.
The pattern used looks a little like the lower half of Cinco de Mayo skull designs, but with bright colors against the skin tone.
The researchers’ approach is described as a “gradient-based optimization process to create a universal perturbation (and mask).” This would mean that anybody could wear the same pattern.
They recorded that in real-world experiments with CCTV, the facial recognition was only able to match 3.34 percent of participants wearing the adversarial match, compared to a minimum of 83.34 percent of participants wearing other evaluated masks.
The researchers noted that multiple styles of patterns may be needed to overcome facial recognition models. Tailor-made masks are a dangerous possibility.
Models need to be trained to recognize adversarial patterns as they evolve like all other threats.