# Bill Swearingen Demonstrates Adversarial Detection Pattern > Source: > Published: 2026-08-12 23:22:09+00:00 # Bill Swearingen Demonstrates Adversarial Detection Pattern At Defcon 34, security researcher Bill Swearingen held a black-and-white patterned panel in front of a live person-detection system, lowering its confidence score from above 0.75 to 0.21 and triggering a "No person detected" result. CNET reports that the demonstration involved person detection, not face matching or identity recognition, and did not establish that similar patterns on shirts can reliably defeat surveillance systems. At Defcon 34, security researcher Bill Swearingen demonstrated a black-and-white adversarial pattern that caused a live AI person-detection system to stop classifying him as a person above its configured confidence threshold. CNET reports that the system's score fell from above 0.75 to 0.21 after Swearingen held up the flat panel, producing a "No person detected" result. The distinction between the demonstrated task and facial recognition is central. As CNET explains, surveillance workflows can include person detection, face detection, and facial recognition as separate stages. The Defcon demonstration concerned the first task: whether the system identified a human body in the camera frame. It did not demonstrate that a system failed to find a face or failed to match a face against an identity database. ### From flat panel to clothing According to CNET, Swearingen has spent roughly a year searching for patterns that confuse computer-vision systems used to identify people, and he is putting related patterns on shirts and hoodies. CNET characterizes the effectiveness of worn clothing as the major unproven test. That gap matters technically. A flat panel can be positioned deliberately in front of a camera, while a garment changes shape with body pose, folds, movement, viewing angle, distance, and illumination. In physical adversarial-machine-learning research, these transformations are often the difference between a pattern that affects a single camera setup and one that transfers across real-world conditions. Swearingen's Defcon presentation, titled noRECOGNITION, places the work in a longer lineage of adversarial fashion experiments, including Adam Harvey's CV Dazzle and HyperFace projects and Kate Rose's adversarial license-plate work. The presentation argues that camera systems parse visual inputs and can be influenced when a person controls those inputs. ### What the demonstration establishes The live result is evidence that the tested detector's confidence can be reduced under the specific demonstration conditions. It is not, on its own, a benchmark of robustness across detectors, cameras, poses, or clothing designs. CNET's reporting does not identify the model, training data, camera configuration, or evaluation protocol used in the stage demonstration. For computer-vision teams, the episode illustrates a familiar robustness problem: a classifier threshold can create a binary operational outcome from a continuous confidence score. Teams evaluating physical-world detection systems commonly test adversarial resilience across viewpoints, lighting conditions, occlusions, apparel deformation, and multiple model families before drawing conclusions about deployment risk. ## Key Points - 1CNET's live demonstration showed a visual pattern reducing a person detector's confidence below its configured operational threshold. - 2The result concerns person detection, not identity matching, so it does not establish evasion of a full facial-recognition pipeline. - 3Physical adversarial robustness generally depends on viewpoint, pose, lighting, and garment deformation, variables not established by flat-panel demonstrations. ## Scoring Rationale The demonstration is a useful, timely illustration of physical adversarial risk in computer vision and surveillance pipelines. Its practitioner impact is limited by the absence of a disclosed model, evaluation protocol, and evidence that a wearable pattern transfers across real-world conditions. ## Sources Primary source and supporting public references used for this report. Practice with real Ad Tech data 90 SQL & Python problems · 15 industry datasets [Active Search Campaigns by BudgetEasy](/problems/sql/active-search-campaigns-by-budget) [High CPC Clicks & Poor Landing PagesMedium](/problems/sql/high-cpc-clicks-poor-landing-page) [Campaign ROAS by Attribution ModelHard](/problems/sql/campaign-roas-by-attribution-model) 250 free problems · No credit card [See all Ad Tech problems](/problems/datasets/adtech)