人的臉型和指紋一樣具有明顯特征，所以越來(lái)越多的組織——從警察局到學(xué)校到沃爾瑪（Wal-Mart）——都在借助人臉識(shí)別軟件來(lái)甄別線上照片和真實(shí)場(chǎng)景中的人。

然而人臉識(shí)別技術(shù)也開(kāi)始給個(gè)人隱私帶來(lái)很大的威脅，一些研究者開(kāi)始探求反制措施，其中包括多倫多大學(xué)計(jì)算機(jī)工程系的學(xué)生喬伊·博斯。

博斯稱他開(kāi)發(fā)了一種工具，可以在照片上傳到網(wǎng)絡(luò)前加入額外的元素，從而“打亂”人臉識(shí)別系統(tǒng)。用肉眼看，處理前后的照片沒(méi)有任何區(qū)別，但處理后的照片有隱藏了的特性，可以阻止檢測(cè)系統(tǒng)的運(yùn)行。

博斯說(shuō)：“這一工具可以給人臉圖像添加特殊制作的雜質(zhì)，用于干擾人臉識(shí)別軟件，就像Instagram（圖片和視頻分享社交軟件——譯注）的過(guò)濾器。”

博斯告訴《財(cái)富》雜志，這個(gè)工具很快會(huì)以手機(jī)應(yīng)用或者瀏覽器插件的形式出現(xiàn)，他也已經(jīng)在GitHub（提供軟件源代碼托管服務(wù)網(wǎng)站——譯注）的個(gè)人主頁(yè)上分享了底層代碼。

在這個(gè)技術(shù)猖獗的時(shí)代，阻撓人臉識(shí)別系統(tǒng)運(yùn)作的反制措施，或許會(huì)受到很多人的歡迎。

有些私企會(huì)動(dòng)腦筋，比如抓住人們對(duì)校園槍擊案的焦慮，向校區(qū)推廣人臉檢測(cè)系統(tǒng)，盡管懷疑人士批駁這種系統(tǒng)反而造成了“安全威脅”，而且也不能阻止校園槍擊案的再次發(fā)生。而據(jù)《福布斯》（Forbes）上周的報(bào)道，亞馬遜網(wǎng)絡(luò)服務(wù)（ Amazon Web Services）正在向顧客售賣(mài)人臉識(shí)別技術(shù)，價(jià)格竟低至10美元。

這樣一來(lái)，博斯所開(kāi)發(fā)的工具，通過(guò)減少人臉檢測(cè)軟件所需的有效人臉數(shù)量，就能減緩人臉識(shí)別技術(shù)的大范圍使用。

不過(guò)博斯的工具是預(yù)防性的，對(duì)于那些已經(jīng)獲取真實(shí)人臉信息，或者使用探頭在真實(shí)世界監(jiān)控人們的公司，這一工具就無(wú)能為力了。博斯認(rèn)為，要預(yù)防此類的人臉識(shí)別，人們也可以使用一些技巧，比如帶特殊形狀的眼鏡以糊弄檢測(cè)系統(tǒng)，甚至在臉上貼點(diǎn)小標(biāo)簽。

貓鼠游戲

目前博斯的工具只對(duì)某幾類的人臉識(shí)別軟件有效，說(shuō)得具體點(diǎn)，就是它能破壞那些訓(xùn)練模式（用于訓(xùn)練軟件所設(shè)置的機(jī)器學(xué)習(xí)數(shù)據(jù)庫(kù)）可公開(kāi)獲取的軟件。

有些安全軟件公司銷售的人臉檢測(cè)系統(tǒng)使用公開(kāi)的數(shù)據(jù)庫(kù)，而有些則不是，最有名的是Facebook, 他們擁有自主產(chǎn)權(quán)的人臉檢測(cè)系統(tǒng)，博斯的工具對(duì)它就不能奏效。

博斯懷疑，F(xiàn)acebook可能是在綜合使用不同的人臉識(shí)別技術(shù)，以對(duì)付各種反制措施，比如他所開(kāi)發(fā)的工具。但博斯認(rèn)為他的人臉識(shí)別破壞軟件最終也能阻撓Facebook，從而開(kāi)啟一場(chǎng)識(shí)別人臉技術(shù)的開(kāi)發(fā)者和掩藏人臉技術(shù)的開(kāi)發(fā)者之間的貓鼠游戲。

這也提出了一個(gè)問(wèn)題，阻撓人臉識(shí)別的工具會(huì)不會(huì)被商業(yè)化。博斯說(shuō)已有幾家風(fēng)投找過(guò)他，但他決定暫不參與，今年秋季他會(huì)繼續(xù)他在麥吉爾大學(xué)的博士生研究項(xiàng)目。（財(cái)富中文網(wǎng)）

譯者：Hank?

The shape of your face is as distinct as your fingerprint. That’s why a growing number of organizations — from police forces to schools to Wal-Mart — are using facial recognition software to identify you in online photos and in real world locations.

But facial recognition technology is beginning to pose a major privacy threat, which has led researchers to explore ways to counteract it. One of them is Joey Bose, a computer engineering student at the University of Toronto.

Bose claims he has developed a tool to “break” facial recognition systems by adding extra elements to photos before they are uploaded to the Internet. The photos don’t look any different to the naked eye, but the hidden features thwart detection systems.

“It adds specially-crafted noise for the face images. It’s trained to attack facial recognition software,” he said. “Think of it as Instagram filter.”

Bose told Fortune the tool will soon be available as a phone app or plug-in for web browsers, and that he has shared the underlying code on his GitHub page.

This opportunity to thwart facial recognition will likely be welcome by many people at a time when the technology is becoming more pervasive.

Private companies, for instance, have seized on anxiety over school shootings to sell face-detection systems to school districts — even as skeptics pan this as “security theater” that’s unlikely to prevent more shootings. Meanwhile, Forbes last week reported that Amazon Web Services is selling facial recognition technology to all comers for as little as $10.

Bose’s could thus slow the spread of the technology by reducing the number of faces available to companies that make the detection software.

His tool, however, is only a pre-emptive measure and does not address situations where a company already has an image of someone’s face and uses a camera to detect them in the real world. In order to prevent this sort of recognition, Bose says, people can employ tactics like wearing glasses with special patterns that fool the detection mechanisms or even put small stickers on their face.

A Cat-and-Mouse Game

For now, Bose’s tool only works to thwart certain types of facial recognition software. Specifically, it can break the software if the training model — the machine learning data set used to train the software — is publicly available.

While a number of facial detection systems sold by security companies rely on these publicly available data sets, other companies, notably Facebook, have their own proprietary versions that Bose’s tool can’t defeat.

Bose suspects that Facebook uses an ensemble of different facial recognition techniques in order to overcome counter-measures, like the one he developed, to fool its software. But he predicts his facial recognition duping tool will eventually be able to thwart Facebook, and set off a cat-and-mouse game between developers seeking to detect faces and those seeking to disguise them.

This raises the question of whether companies will seek to commercialize tools that thwart facial recognition. Bose says he has already been approached by a number of venture capitalists, but that he’s decided to pass for now. Instead, he says he plans to continue his research in a PhD program at McGill University starting this fall.