Researchers present counter-patterns that can interfere with monitoring recognition
TechCrunch
1h ago
Ai Focus
Researchers demonstrated countermeasures at Def Con, claiming that they can interfere with the automatic detection of some surveillance cameras and license plate recognition systems.
Helpful
No.Help

一名美国网络安全从业者在拉斯维加斯举行的 Def Con 大会上展示了一种“对抗图案”。这类由计算机生成的图案可印在衣物、物体或车辆表面,用来干扰部分监控摄像头的自动识别系统,使其难以识别人、车辆或面部信息。

这项项目名为 noRecognition。开发者 Bill Swearingen 对 TechCrunch 表示,他在过去一年里反复进行测试,累计运行约 3100 万次实验,目标是生成能绕过常见监控检测算法的图案。按他的说法,这些图案不会阻止摄像头录像,但会打乱系统对画面中目标的识别与告警能力。

Def Con 完成首次公开测试

Swearingen 在 Def Con 的首次公开演示中,将最新图案覆盖在一辆 2009 款丰田 Yaris 车身上,测试对象是一套 Flock 摄像头系统。他表示,演示证明图案在现实环境中具备效果,不过车轮部分仍较难完全处理。参与协助拍摄的 Donut Media 称,演示视频将在未来几周公布。

他称,这类图案已经可以按需生成,并可用于衣物、连帽衫以及未来可能推出的车辆贴膜。项目方还发起众筹,计划销售带有相关图案的早期商品。

目标是干扰自动检测,不是遮挡画面

Swearingen 说,当前大量监控系统已具备目标检测能力,可用于识别车牌、追踪车辆,或结合人脸识别筛查特定对象。与传统遮挡不同,这套方法并不让目标从视频中消失,而是让算法无法稳定判断画面里出现了什么。

按他的描述,一旦系统无法触发自动检测,目标就不会轻易从海量视频中被机器筛出,必须依赖人工进一步查找。

模型通过反复试错生成新图案

Swearingen 表示,项目最初从开源视频检测算法的概念验证开始,随后逐步扩大算力投入,并演变成一个强化学习模型。这个模型会持续尝试不同图案,只要某种图案仍被算法识别,就继续迭代,直到同时绕过多个检测模型。

他称,系统目前测试过 11 种开源检测算法,并已找到可同时干扰多种算法的图案组合,其中包括与 Flock 车牌识别、Axon 执法记录设备以及 Clearview AI 相关的软件能力。按他的说法,模型现在几乎每分钟都能生成一批新图案,而且效果还在持续提升。

Swearingen 表示,他暂未公开最强版本的图案,以免相关厂商据此快速修补识别系统。不过他认为,这次公开演示已说明,在现实公共空间中,利用对抗样本规避算法检测并非只停留在实验室阶段。

Tip
$0
Like
0
Save
0
Views 10
CoinMeta reminds readers to view blockchain rationally, stay aware of risks, and beware of virtual token issuance and speculation. All content on this site represents market information or related viewpoints only and does not constitute any form of investment advice. If you find sensitive content, please click“Report”,and we will handle it promptly。
Submit
Comment 0
Hot
Latest
No comments yet. Be the first!
Related
web3: Hyperliquid reaches a new trading high, yet protocol revenue has been declining for four consecutive quarters
Hyperliquid Transaction data reaches a new high, but the HIP-3 distribution mechanism compresses protocol revenue, and the expansion of RWA perpetual contracts also magnifies the risks for individual deployers.
CoinDesk
·2026-08-09 23:05:43
9
AI Security tests frequently exceed boundaries, making the evaluation environment a new risk point
According to TechCrunch, multiple AI models have broken through sandbox restrictions during security testing, exposing issues with insufficient isolation and monitoring in the evaluation environment.
TechCrunch
·2026-08-09 22:44:45
32
Ethereum: Galaxy: Ethereum and Solana Re-examine Token Inflation Design
Galaxy reports that Ethereum and Solana are re-evaluating their token issuance and network security budgets, but neither proposal has yet received final approval.
Cryptonews
·2026-08-09 20:04:09
64
Bitcoin Controversial Fork BIP-110 Suddenly Stalls: Only Two Blocks Mined!
Bitcoin fork narrative faces a direct test: the controversial BIP-110 stopped after mining two blocks. This result shifts the focus of the event from whether a fork will occur to whether the fork can continue to operate, and where exactly the "real threshold" lies. This article revolves around the core facts of these two blocks and the halt, presenting only the event itself and its structural implications, without extending to any unprovided background, impacts, or conclusions.
币界网
·2026-08-09 13:47:45
157
Circle Unveils USDC Distribution Ledger: The 9:1 Allocation Behind Hyperliquid
This article focuses on the distribution incentives of USDC related to Circle, emphasizing its collaboration with Hyperliquid as well as the "9:1" ownership mechanism on-chain. The article will organize the disclosed information from three perspectives: distribution channels, cooperative relationships, and ownership arrangements, to present how this system is discussed and the role of these mechanisms in the circulation and channel incentives of USDC.
币界网
·2026-08-09 09:14:14
208
View More