China Uyghur Analytic Projects Require Intel And NVIDIA, Intel Condemns, NVIDIA SilentBy: IPVM Team, Published on Dec 02, 2019
At least 8 PRC China police projects require NVIDIA and Intel chips to power their Uyghur-detecting analytics, according to procurement documents found by IPVM.
Intel promptly condemned the usage while NVIDIA remains silent to IPVM inquiries. Intel told IPVM it had no idea about this and condemned human rights abusing applications of their products, adding that it did not always have control over where its chips end up. NVIDIA, which has heavily promoted Tesla GPUs for PRC smart cities in the past, did not respond to repeated requests for comment from IPVM.
Meanwhile, the PRC government has increasingly promoted the use of its own domestic chips, in sharp contrast to the ongoing requirements for US chips in these video surveillance projects.
In this note, we examine this news and the human rights concerns, including:
- Background: Ethnicity Analytics in the PRC
- Evidence Of Ethnicity Analytics Mandated from the Top
- Examples: 8 Separate Police Projects
- Western Equipment Easily Available
- Not Subject to Sanctions
- Intel: "Opposes Any Abuse of Human Rights", Can't Always Control Where Chips End Up
- NVIDIA No Response, Previously Touted Tesla GPUs for China Smart Cities
- Amnesty International: "Extra Precautions" Needed for AI Equipment
Background: Ethnicity Analytics in the PRC
IPVM found government surveillance requirements and at least a dozen projects requiring analytics to identify Uyghurs across the country.
Eight Specifying Intel and NVIDIA
Of the 12 found, eight separate PRC police projects required NVIDIA/Intel equipment to power their systems, particularly NVIDIA's Tesla series GPUs. All of the projects were relatively recent, from 2018 and 2019.
The Public Security Bureau (police department) of Jinyun County in Zhejiang province issued an equipment request for a video surveillance network which mandated Uyghur analytics and called for an NVIDIA Tesla GPU and Intel chips:
Translated into English, this spec reads (emphasis added):
Support historic statistics on people’s age, sex, whether they have glasses, whether they have hair bangs, whether they have sunglasses, and whether they are Uyghurs.
For further context, here is the full spec translated into English:
Lankao County in Henan Province issued a procurement document for a video surveillance network which includes a "live face comparison system" allowing police to search for Uyghurs:
★the system supports filtering functions for both real-time and historic human image captures. Filter criteria includes time, location, human features (such as age range, ethnicity (Uyghur), mask, bangs, sex, glasses, etc).
This particular system requires eight Nvidia Tesla P4 GPUs and two Intel E5 2680 CPUs:
A Gansu Provincial Government tender for a surveillance network includes a "high-performance face recognition analysis system" with Uyghur analytics:
(15): Supports automatic labeling of captured faces [...] including age, gender, ethnicity (Han and Uyghur), color of clothes and pants, movement direction, whether or not [wearing] a backpack, whether using an umbrella;
This system requires Nvidia Tesla GPUs and Intel CPUs:
Yongtai County in Fujian province issued documents for a surveillance network for its PSB which mandates Uyghur analytics:
In order to help the police more effectively screen targets, the system should be able to model each passerby and classify key features. The police investigators can use a number of attribute classifications to screen passers-by, such as age, gender, whether they are wearing glasses, or whether they are Uyghurs, so as to quickly locate the profile information of the suspect.
The network requires NVIDIA Tesla GPUs and Intel chips:
Yanghe New District
Yanghe New District in Jiangsu Province issued a call for an 'intelligent community' project built by the local PSB which mandates Uyghur analytics:
Live portrait comparison system: The system supports face detection function: support attribute recognition of captured passers-by; support analysis of passerby age, age range, ethnicity attributes, gender, whether wearing glasses, or Uyghurs;
NVIDIA Tesla P4 GPUs and Intel chips are requested for the project's servers:
Jingbian County in Shaanxi province called for equipment for its local PSB, including an "intelligent portrait capture machine" with Uyghur analytics:
★ Uygur/Non-Uyghur attribute: It should support the detection of Uyghur/non-Uyghur facial attributes
The "machine" (likely a server, although this is not explicitly stated) requires Intel and NVIDIA chips:
Zibo City's Public Security Bureau (Linyi Branch) issued a call for equipment for its video analysis platform, which mandates Uyghur analytics:
Ethnicity: should support the identification of Han, Uighur, other
On the project's equipment list are Intel chips and GPUs with NVIDIA Pascal architecture:
Chongqing Municipality issued a tender for a "public security video surveillance network" in Nan'an district which included Intel chips for a server capable of detecting Uyghurs:
The underlined Chinese text reads "whether they are Uyghurs"; see below for a fuller translation:
Supports control of key persons [...] feature classification of captured pictures, and search according to criteria including: gender, age, whether [they have] glasses, sunglasses, whether [they have] bangs, whether they are Uyghurs
"Key persons" is a commonly-used term in PRC security which refers to Uyghurs and other groups that police believe are "suspected of threatening national security or public order", as reported in Foreign Policy magazine.
Later in the document, NVIDIA Tesla GPUs/Intel CPUs are requested for a server capable of "feature extraction":
The PSB of Huaiyuan County in Anhui Province called for a "portrait recognition system" which mandates Uyghur analytics:
Personnel national identification: It has the function of identifying the ethnic groups (including Han and Uygur) in the uploaded images
The system's servers use Intel CPUs and a NVIDIA GTX1080 GPU:
Western Equipment Widely Available
Intel CPUs and NVIDIA GPUs are widely available almost anywhere in the world. For example, NVIDIA's Tesla GPUs are easy to purchase on Alibaba (China's equivalent to Amazon):
For this reason, Intel and NVIDIA chips being used for Uyghur analytics is not comparable to Hikvision or Dahua, who have directly won over $1 billion in video surveillance projects in Xinjiang.
Not Subject to Sanctions
The US government imposed sanctions on Hikvision and Dahua last month for their role in Xinjiang surveillance. However, a Tesla GPU being used by PRC police for Uyghur detection is not illegal/sanctionable, as long as the police are not in Xinjiang itself.
This is because the US sanctions only cover supplying US technology to specific Chinese companies and police departments in Xinjiang itself. However, the PRC's policies against Uyghurs extend far beyond Xinjiang, since many Uyghurs have migrated within the PRC - hence the use of Uyghur analytics by police all over the country.
Intel: "Opposes Any Abuse of Human Rights", Can't Always Control Where Chips End Up
Intel sent IPVM the following statement, saying "we do not always know nor can we control" the end-use of their equipment:
Intel opposes any abuse of human rights and has a policy to not ship any products in situations such as these unless we have high confidence that our products are not being used in connection with human rights abuses. Intel processors are general-purpose computing products that can be incorporated into countless systems and applications, and Intel products are resold through a large number of distributors and OEMs. We do not always know nor can we control what products our customers create or the applications end-users may develop. Notably, the RFP’s you cite are directed to others, not Intel. [emphasis added]
NVIDIA No Response, Previously Touted Tesla GPUs for PRC Smart Cities
NVIDIA did not respond to repeated requests for comment from IPVM. In a 2017 blog post about its Metropolis smart city program, NVIDIA promoted the use of Tesla GPUs in China, with Hikvision even praising their "outstanding performance":
[...] Offering a striking demonstration of what Metropolis makes possible, Hikvision has achieved recall rates of more than 90 percent for its identification and matching technology despite data volumes growing by more than 10x. This makes it easier to find lost people in crowded places. To achieve this, Hikvision combines a camera and network video recorder with NVIDIA Jetson at the edge, cloud servers powered by NVIDIA Tesla P4 GPU accelerators
“The GPU’s outstanding performance and NVIDIA’s end-to-end AI and deep learning platform can be applied to video streams to create smarter applications for a variety of industries,” said Shiliang Pu, president at Hikvision Research Institute.
Meanwhile, Huawei’s video content management (VCM) product is equipped with Tesla P4 GPU accelerators, improving overall performance by 22x.
To be clear, this September 2017 blog post predates the emergence of Xinjiang as an international human rights crisis and the US sanctions on Huawei/Hikvision/etc.
Amnesty International: "Extra Precautions" Needed for AI Equipment
Joe Westby, a researcher for Amnesty International's technology division, told IPVM that companies/governments need to take "extra precautions" so that AI equipment isn't used to fuel repression in the PRC given "widespread ethnic discrimination" there:
Governments and multinational technology companies must both take extra precautions to address the high risk that AI technology will fuel serious human rights abuses in Xinjiang Uyghur Autonomous Region or elsewhere in China. States must ensure that exports of relevant advanced technology are properly scrutinized for human rights risks. Companies have a responsibility to respect all human rights wherever they operate in the world, but given the extreme state of surveillance in Xinjiang, and widespread ethnic discrimination across China, companies should be on notice to ensure that their technology doesn’t contribute to these abuses.
The challenge, though, is how this can be done given how widely available these chips are.
PRC Chips Instead of US Ones?
One irony of the widespread requirement of US chips, such as Intel and NVIDIA, in these recent surveillance projects, is that the PRC government has been advocating for the use of domestic semiconductors, instead of US ones.
Indeed, AI / robotics expert Noel Sharkey told IPVM:
These [Intel, NVIDIA] are definitely the most powerful chips around for deep learning on the general market but I really think that if they were stopped, China would just find others and I imagine that there are Chines companies already working on developing their own versions.
It is reasonable that PRC companies are but it is unclear how long or how successful PRC companies will be. By contrast, the WSJ recently reported how Huawei is managing to make smartphones without US chips.
The fact that PRC police are using US technology to power their Uyghur analytics leads to several conclusions:
- Once in the PRC, Western-made AI hardware and other technology can have explicitly racist/authoritarian end-uses, regardless of the intent/desire of the original manufacturer.
- Despite the PRC's repeated claims of tech self-sufficiency, it remains heavily dependent on US chips & tech, even in sensitive police projects. The Gansu Province project examined by IPVM even mandated "completely domestic" algorithms but not chips.
- While the human rights concerns here are obvious, placing controls on such equipment would be difficult. NVIDIA GPUs originated as graphics cards for gaming and remain very easy to obtain to virtually everyone, even if the end use is Uyghur detection and PRC surveillance. Intel processors, of course, are also incredibly common and easy to get all over the world.
The rise of AI video surveillance is increasing ethical issues and questions for both technology developers and governments worldwide.