ACLU并没有错：某些技术确实可以促进这种反乌托邦的愿景。 (The ACLU isn’t wrong: Some technology could indeed contribute to this dystopian vision.)

A dystopia is a community or society that is undesirable or frightening — translated as “a bad place.” Some systems attempt to use technology that’s not yet advanced enough, raising the specter of misuse; others invade the privacy of citizens, all for minimal security benefits.

反乌托邦是一个不受欢迎或令人恐惧的社区或社会，被翻译为“不好的地方”。 一些系统尝试使用尚未足够先进的技术，从而增加了滥用的可能性。 其他人则侵犯了公民的隐私，所有这些都是为了获得最小的安全利益。

但是，人工智能视频监控的途径可以在不牺牲公民自由的情况下带来更大的公共安全。 (But there are avenues of AI video surveillance that can bring greater public safety without sacrificing civil liberties.)

The key is using AI to enable human security professionals to do their jobs better, not overextending AI’s capabilities to take over their jobs entirely. I see AI surveillance existing within three main categories: behavioral analysis, facial recognition, and object detection. The first two categories raise concerns. To understand what makes the last more viable, it’s important to break down the issues with the others.

关键是使用AI可以使人类安全专业人员更好地完成工作，而不是过度扩展AI完全接管其工作的能力。 我看到AI监视存在三个主要类别：行为分析， 面部识别和对象检测。 前两类引起关注。 要了解什么使最后一个更可行，重要的是与其他人一起分解问题。

人工智能还不足以进行行为分析 (AI Isn’t Advanced Enough for Behavioral Analysis)

Behavioral analysis is essentially an attempt to detect so-called suspicious behavior before any crimes are committed. The ACLU’s “Th­e Dawn of Robot Surveillance” report touches on a few areas here: human action recognition, anomaly detection, and contextual understanding. Contextual understanding is the most advanced, but researchers have yet to make it genuinely feasible, and it’s unclear whether current technology provides a path to this type of general intelligence.

行为分析本质上是一种在犯罪之前检测出可疑行为的尝试。 ACLU的“ 机器人监视的曙光 ”报告涉及以下几个方面：人体动作识别，异常检测和上下文理解。 上下文理解是最先进的方法，但是研究人员尚未真正使它可行 ，并且尚不清楚当前的技术是否提供了通往这种通用情报的途径。

The problem is that computers lack the “common sense” to relate things to the rest of the world. A computer can recognize a dog thanks to thousands of images of dogs. It’s seen, but it can’t understand the context around a dog.

问题在于计算机缺乏将事物与世界其他地方联系起来的“ 常识 ”。 多亏成千上万的狗的图像，计算机才能识别狗。 可以看到，但无法理解狗的周围环境。

AI无法推断-它只能识别模式。 (AI can’t infer — it can only recognize patterns.)

Some companies are already putting this technology into action, however. The New York Police Department has partnered with Microsoft to produce what it calls the Domain Awareness System. Part of the system would involve smart cameras that aim to detect suspicious behaviors.

但是，一些公司已经将该技术付诸实践。 纽约警察局已经与Microsoft合作生产了所谓的域感知系统。 该系统的一部分将涉及旨在检测可疑行为的智能相机。

But it’s not a finished product; developers have been working with officials to tweak the software since it was put in place. I’m confident that Microsoft will eventually crack the code to getting this technology working, but behavioral detection is still in beta.

但这不是成品。 自该软件推出以来，开发人员一直在与官员合作调整该软件。 我有信心Microsoft最终会破解使该技术起作用的代码，但是行为检测仍处于测试阶段。

行为分析可能可行的一个领域是检测盗窃，这主要是由于缺乏其他可行的选择。 (The one area where behavioral analysis may be feasible is in detecting theft, and that’s primarily due to the lack of other viable options.)

Without Amazon Go-style camera installations, tracking every item in a store isn’t possible — so the next-best option is to guess whether a person is suspicious based on certain detected behaviors. But that, in itself, draws civil liberties concerns. The ACLU report notes the problems inherent in identifying “anomalous” behavior and people.

如果没有安装Amazon Go风格的摄像头，就不可能跟踪商店中的所有物品-因此，下一个最佳选择是根据检测到的某些行为来猜测某人是否可疑。 但这本身引起了公民自由的关注。 ACLU报告指出了识别“异常”行为和人员时固有的问题。

The big concern with action-detecting systems, then, is that the technology isn’t yet advanced enough to produce accurate results outside of small niches such as theft. Security professionals would be left with ambiguous data to interpret — likely reinforcing existing biases rather than making objective observations.

因此，动作检测系统最关心的是该技术还不够先进，无法在盗窃等小众市场之外产生准确的结果。 安全专业人员将留有模棱两可的数据来解释-可能会加剧现有偏差，而不是进行客观观察。

面部识别为不良演员创造目标 (Facial Recognition Creates a Target for Bad Actors)

Facial recognition works significantly better than behavioral analysis, and many law enforcement agencies and security firms already use the technology. That’s not to say it’s anywhere near perfect. False positives are — especially when it comes to people of color. As a result, cities like San Francisco have banned the use of facial recognition software by local government agencies.

面部识别比行为分析要好得多，而且许多执法机构和安全公司已经在使用该技术。 这并不是说它接近完美。 误报是-尤其是对于有色人种 。 结果，像旧金山这样的城市已禁止地方政府机构使用面部识别软件。

即使面部识别技术100％准确，也可能无法阻止最严重的犯罪。 (Even if facial recognition technology was 100% accurate, it still might not stop the worst of crimes.)

Acts of violence such as mass shootings are regularly perpetrated by students, family members, employees, or customers: in other words, people who “” in the location. It’s unlikely that a facial recognition system would flag these individuals.

学生，家庭成员，雇员或顾客会定期进行大规模射击等暴力行为：换句话说，就是在该地点“打”的人。 面部识别系统不太可能会标记这些人。

保护面部周围居民和家中的人真的只是面部识别的“错误选择”吗？ (Is it really just a “false choice” for facial recognition to protect people in their own neighborhoods and homes?)

Then, there’s the severe invasion of privacy that’s required to make this technology work. To identify a suspect, facial recognition requires an external database of faces as well as personally-identifying information to match the face to a name. A database like that is an attractive target for bad actors.

然后，要使这项技术正常工作，就必须严重侵犯隐私 。 为了识别嫌疑人，面部识别需要外部人脸数据库以及用于将人脸与姓名匹配的个人识别信息。 像这样的数据库对于不良行为者来说是一个有吸引力的目标。

Case in point: In July 2019, the U.S. Customs and Border Protection Agency announced that hackers had gained access to a third-party database containing license plate numbers and photo IDs. The CBP has, in recent years, begun collecting facial recognition data and fingerprints, among other things, from foreign travelers. It’s not hard to imagine a world where hackers gain access to this kind of information, endangering individuals as a result.

案例分析：2019年7月， 美国海关与边境保护局宣布黑客已经可以访问包含牌照号码和带有照片ID的第三方数据库。 近年来，CBP已开始从外国旅行者那里收集面部识别数据和指纹。 不难想象，世界上黑客可以访问此类信息，从而威胁到个人 。

我们不必为了安全而牺牲隐私 (We Don’t Have to Sacrifice Privacy for Safety)

The issues associated with behavioral analysis and facial recognition all lead back to the human element of threat detection. That’s where object detection differs. Object detection, as its name suggests, is entirely self-detained and functions by flagging known items, not individuals. Therefore, people’s private information remains just that — private.

与行为分析和面部识别相关的问题都可以追溯到威胁检测的人为因素。 那是物体检测的不同之处。 顾名思义，对象检测完全是自持的，并且通过标记已知项目 (而不是个人)来起作用 。 因此，人们的私人信息仍然只是私人信息。

Still, there’s room for improvement. For instance, nonthreatening but unusual items such as power tools may be flagged. And the technology cannot detect concealed weapons that aren’t otherwise raising suspicions.

尽管如此，仍有改进的余地 。 例如，可能会标出非威胁性但异常的物品，例如电动工具。 而且该技术无法检测出没有引起怀疑的隐藏武器。

We’re amid constant advancement and refinement. But currently, object detection — flaws and all — is the best avenue for keeping citizens safe without compromising their privacy. As AI video surveillance continues to advance, we should focus on letting security staff members do their jobs better instead of trying to automate them away.

我们正处在不断进步和完善中。 但是目前，物体检测(包括缺陷和所有缺陷)是确保公民安全而不损害其隐私的最佳途径。 随着AI视频监控的不断发展，我们应该集中精力让安全人员更好地完成工作，而不是试图使他们自动化 。

During most mass shootings, for example, police officers know little about the shooter’s location, appearance, or armaments. Not having real-time updates limits police in their ability to respond.

例如，在大多数大规模枪击事件中，警官对射手的位置，外观或武器一无所知。 没有实时更新限制了警察的响应能力。

但是，物体检测AI监视系统可以高精度地检测武器，遗弃物体和其他可能威胁的物品。 (Object detection AI surveillance systems, however, can detect weapons, abandoned objects, and other possibly threatening items with high accuracy.)

Upon detection, the AI system can then notify security professionals of their location in real-time. Real-time allows for nearly instant responses. In the recent Virginia Beach shooting, officers took nearly 10 minutes to locate the gunman after they entered the building, and by the time the officers had subdued the gunman, the attack had lasted about 40 minutes.

一旦检测到，AI系统便可以实时通知安全专业人员其位置。 实时允许几乎即时的响应。 在最近的弗吉尼亚海滩枪击事件中 ，警官进入建筑物后花了将近10分钟的时间来找到枪手，而当警察制服枪手时，袭击持续了大约40分钟。

That may not seem like long, but when there’s an active shooter involved, every second counts. An active shooter is the exact sort of practical situation in which AI can offer real, valuable data instead of a false sense of security.

这似乎并不长，但是当涉及到一个活跃的射手时，每一秒都很重要。 主动射击是一种实际情况，在这种情况下，AI可以提供真实，有价值的数据，而不是错误的安全感。

个人隐私和公共安全不必相互排斥，并且在隐私和安全之间进行选择可能不是错误的选择。 (Personal privacy and public safety don’t have to be mutually exclusive and may not be a false choice between privacy and safety.)

It’s time we stopped trying to overextend AI video surveillance to do the work of security professionals. Instead, let’s focus on technology that already works to provide them with real, actionable data that can help them do their jobs better. We can create a safer society without sacrificing civil liberties.

是时候我们停止尝试过度扩展AI视频监控来完成安全专业人员的工作了。 相反，让我们关注已经为他们提供真实，可行的数据的技术，这些数据可以帮助他们更好地完成工作。 我们可以在不牺牲公民自由的情况下建立更安全的社会。

Image credit: rishabh-varshney — Unsplash

图片来源：rishabh-varshney — Unsplash

The False Choice Between Privacy and Safety in Smart Surveillance was originally published on ReadWrite on October 16, 2019 by Ben Ziomek.

智能监视中隐私与安全之间的错误选择 最初 由 Ben Ziomek 于2019年10月16日 发布在 ReadWrite 上 。