A new research cooperation between Israel and Japan believes that the pedestrian testing system has inherent weaknesses, so that well -known individuals can escape the facial recognition system through the most impossible field of monitoring networks through a carefully planned route.
With the help of public videos of New York and San Francisco, researchers have developed a method of automation method based on the most popular object recognition system that may be used in public networks to calculate such ways.
Three cross -ports used in the study: Shibuya Crossing, Tokyo, Japan; Broadway in New York; and Castro, San Francisco. Data source: https: //arxiv.org/pdf/2501.15653
Through this method, it can be generated Hot picture This is divided into the area in the camera feed, which makes pedestrians most unlikely to provide positive facial recognition strikes:
On the right, we see the confidence diagram generated by the researcher method. The red area indicates that the reliability is low, and the pose that may hinder the facial recognition, the camera posture and other factors configuration.
Theoretically, this method can spread the “identification friendly” path from A to B in any other calculated position to play a role.
The new thesis proposed a title entitled Location -based privacy technology (L-PET); Location -based adaptive threshold (L-BAT) Basically run the same routine, and then use information to strengthen and improve surveillance measures, rather than design methods to avoid being recognized; in many cases, if you do not invest further in monitoring infrastructure, you cannot invest in further investment, and you will not be able to invest in further investment. For this kind of improvement.
Therefore, this article has established a potential technological war in this article to find a way to optimize its ways to avoid detection.
The previous method of frustration detection is not so elegant, and confuse the detection algorithm by confusing the algorithm by confrontation (such as TNT attacks) and using printing patterns.
In 2019, the “deception automatic surveillance camera: the confrontation of the attacker detection” shows a confrontation printing mode that can persuade the identification system. No one is found, which allows “an invisible”. Data source: https: //arxiv.org/pdf/1904.08653
Researchers behind the new papers observe that their methods need less preparation, and do not need to design confrontation items (see the figure above).
The title of the paper is entitled A kind of privacy enhancement technology can evade detection through street cameras without having to use confrontation accessoriesFive researchers from Negiv Ben Gurian University and Fules Tong Co., Ltd..
Methods and tests
According to previous works, such as confrontation masks, advhat, confrontation with patch and other similar outings, researchers believe that pedestrians “attackers” know which object detection system is using in the surveillance network. In fact, this is not an unreasonable hypothesis, because this is the most advanced open source system such as Cisco and Ultralytics (currently the core driving force developed by YOLO).
This article also assumes that pedestrians can access the real -time flow on the Internet that is fixed on the calculated position. This is once again a reasonable assumption that most of them may have the intensity of coverage.
Websites such as 511ny.org provide access to many surveillance cameras in New York City. Source: https: //511ny.or
In addition, pedestrians also need to access the methods proposed, as well as the scene itself (that is, to determine the cross and route of the “security” route).
In order to develop the L-PET, the author evaluates the impact of the pedestrian angle relative to the camera. The height of the camera; the effect of the distance; and the influence of the day. In order to obtain the truth of the ground, they shot a person with 0 °, 45 °, 90 °, 135 °, 180 °, 225 °, 270 °, and 315 ° angle.
Researcher’s ground truth observation.
They repeated these changes at the height of three different cameras (0.6m, 1.8m, 24m), and have different lighting conditions (morning, afternoon, “laboratory and” laboratory conditions)).
Feed this lens into faster R-CNN and YOLOV3 object detectors. They found that the confidence of the object depends on the acuity of pedestrian angles, the distance between pedestrians, the height of the camera, and the weather/lighting conditions*.
Then, the author tested a wider range of object detectors in the same situation: faster R-CNN; yolov3; ssd; div
The author pointed out:
“ We found that the architecture of all five object detectors was affected by pedestrian location and ambient light. In addition, we found that the effect continues to exist at all environmental light levels for the five models (YOLOV3, SSD and RTMDET).
In order to expand the scope, researchers used videos shot by traffic cameras from three locations: Shibuya Crossing in Tokyo, Castro Distribution of Broadway and San Francisco in New York.
Each position is provided with five to six recording, and each recording has a video of about four hours. In order to analyze the detection performance, the frame is extracted every two seconds, and the faster R-CNN object detector is used for processing. For each pixel obtained, the method estimates the average confidence of the “person” detection borders in the pixel.
“ We found that in all three positions, the confidence of the object detector depends on the position of the personnel in the framework. For example, in Shibuya Crossing video, there is a lot of confidence, farther from the camera, and close to the camera, where there is a pole part covering the pedestrians.
The L-PET method is essentially the program, which can be said to be “weaponized” to obtain a path through urban areas, which is the most impossible to cause pedestrians to be successfully recognized.
In contrast, the L-BAT follows the same process. The difference is that it updates the score in the detection system, thereby creating a feedback circuit aiming to avoid the L-PET method and make the “blind spots” of the system more effective.
(However, in fact, the coverage of the obtained heat map improving coverage not only needs to be upgraded to the camera at the expected location; according to the test standard, including the location, the additional camera can be installed to cover It can be said that the L-PET method does upgrade this special “cold war” to a very expensive situation)
In the observed Castro Street area, the average pedestrian detection confidence of each pixel is analyzed in various detector frameworks. Each video is recorded under different lighting conditions: sunrise, daytime, sunset and two different night settings. For each lighting scene, the results are displayed.
After the researchers converted the pixel -based matrix to a chart that is suitable for the task, the Dijkstra algorithm was adjusted to calculate the best path to allow pedestrians to browse the surveillance and detection area.
The algorithm has been modified instead of finding the shortest path to minimize the test confidence, and treats high confidence as a “high cost” area. This adaptation enables the algorithm to identify the route through the blind spot or low detection area, thereby effectively guiding pedestrians along the path to visible monitoring system.
Visualization describes the heat map of the scene from a pixel -based matrix to a graph -based representation.
The researchers evaluated the impact of the L-BAT system on pedestrian detection. The data set was based on the data set recorded by the four-hour public pedestrian traffic. To fill the set, use the SSD object detector to deal with the frame every two seconds.
From each frame, select a boundary box, which contains a detected person as a positive sample, and another random area of the person who is not detected is used as negative samples. These twin samples form a data set to evaluate the two faster R-CNN model-one uses L-BAT, one is not.
The performance of the model is evaluated by checking the accuracy of the positive and negative samples: the boundary box of the overlapping front sample is considered a real front, and the negative samples overlapped by the border frame are marked as false positive.
The indicators used to determine the reliability of the L-BAT detection are the area (AUC) under the curve; the true positive rate (TPR); the error report (FPR); the true positive confidence of peace. Researchers asserted that using L-BAT enhanced the test confidence while maintaining a high and true positive rate (although the false positive increase was slightly increased).
At the end of the author, there are some limitations of this method. One is that the hot map generated by the method is particularly at a specific time of the day. Although they did not explain it, this shows that a larger multi -layer method is needed to solve the more flexible deployment time.
They also observed that the heat map will not be transferred to different model architectures and is related to specific object detector models. Because the proposed work is essentially conceptual verification, more Adroit architectures may be developed to make up for this technical debt.
in conclusion
Any new attack method of “paying a new surveillance camera” in this solution has certain advantages, because expanding the citizen’s camera network in a highly residual region may be politically challenging and represent a significant citizen cost. Usually Voters need voters to authorize.
Maybe the biggest problem with this work is “Does the closed source monitoring system use the open source SOTA framework (such as YOLO)?”Essence Of course, it is impossible to know, because manufacturers (at least in the United States) of the ownership of so many state and citizen camera networks will argue that disclosure of this usage may open their attacks.
Nevertheless, the migration of government IT and internal proprietary regulations to global and open source regulations will show that any person who argues to the author (for example) YOLO may immediately trigger the award.
* Generally, I usually provide related table results in the paper, but in this case, the complexity of the thesis table makes them dissatisfied with leisure readers, so the abstract is more useful.
The first publishing was on January 28, 2025, Tuesday, Tuesday
