Amazon's AI Is Helping the Police Watch You

The ACLU has raised concerns over law enforcement agencies' use of Rekognition, a facial and object recognition AI developed by Amazon that can be used for person tracking, among other applications.

The American Civil Liberties Union (ACLU) is not happy after recently learning that Amazon is licensing AI capable of tracking persons of interest in real time to law enforcement agencies. Amazon markets the AI, called Rekognition, for a variety of applications from the innocuous (allowing videos to be searchable by the people and objects that appear in them) to more concerning applications, such as real-time tracking and security. Amazon distributes the deep learning-based system via an API and is constantly training it, meaning all of Amazon's Rekognition customers can benefit as the AI grows smarter.

At the Amazon Web Services conference held in South Korea earlier this month, Ranju Das, General Manager of Amazon Rekognition, said that Rekognition can be set up to monitor footage from traffic cameras and other surveillance and security cameras to identify persons of interest in real time—and even reconstruct their past movements. The system can recognize people even when their face isn't fully visible and is capable of recognizing up to 100 individuals in a single scene.

Ranju Das, General Manager of Amazon Rekognition, discusses the AI during an Amazon Web Services conference in South Korea. 

Amazon already lists several entities as customers including C-Span, Motorola, and Pinterest. But two customers in particular have raised the ire of the ACLU—the city of Orlando and the Washington County Sheriff Office. Both are using Rekognition for law enforcement applications.

In a statement on Amazon's Rekognition website, John Mina, Police Chief for the city of Orlando, said that the city is conducting a pilot program using Rekognition “...that will use existing city resources to provide real-time detection and notification of persons-of-interests, further increasing public safety and operational efficiency opportunities for the city of Orlando and other cities across the nation."

In his own statement, Chris Adzima, Senior Information Systems Analyst at the Washington County Sheriff’s Office, said the sheriff's office has been using Rekognition to comb through thousands of photo records to identify suspects. “Within one week of going live, the solution built on Rekognition identified a suspect for a cold case that led to an arrest through due process,” Adzima said.

Even though the law enforcement organizations have pledged to only use Rekognition in accordance with all applicable laws, the ACLU has issued a formal call demanding that Amazon cease allowing governments to use Rekognition.

In a blog from the ACLU, Matt Cagle, Technology and Civil Liberties Attorney for the ACLU of Northern California, and Nicole Ozer, Technology and Civil Liberties Director for the ACLU of California, called Rekognition “powerful and dangerous” and voiced concerns about how easily it can be used to violate civil liberties:

“Amazon says Rekognition can be used to identify 'people of interest,' raising the possibility that those labeled suspicious by governments—such as undocumented immigrants or Black activists—will be seen as fair game for Rekognition surveillance. It also says Rekognition can monitor 'all faces in group photos, crowded events, and public places, such as airports' at a time when Americans are joining public protests at unprecedented levels.”

In an interview with NPR, Cagle said, “Amazon is handing governments a surveillance system primed for abuse...And that's why we're blowing the whistle right now."

The Technology

Amazon has not made technical details about how Rekognition works publicly available. However, based on the AI's reported functionality, it can be surmised that Rekognition's deep learning is a combination of a convolutional neural network (CNN) and some kind of recurrent neural network (RNN)—most likely a long short-term memory network (LSTM). CNN's are the go-to neural networks for image and video recognition tasks, as they are designed based on computer vision principles and can be trained to classify objects (and people) based on their components. A CNN can learn to recognize eyes, mouths, noses, ears, etc. and know that the combination of these things constitutes a face.

RNNs are a type of neural network with built-in loops that allow for information to be retained, essentially giving AI a memory. LSTMs are a type of RNN that take this functionality even further, allowing the neural network to work with larger data sets. In essence, LSTMs enable AI to remember more past information and use it to infer and predict things based on context.

A CNN/LSTM combination would allow for the type of functionality Amazon describes for Rekognition. Not only would the system be able to recognize an individual, it could also “remember “ that person and find him or her in multiple instances over time (i.e., across many surveillance videos).

In a law enforcement application, this means that the AI, after being trained to recognize a suspect's face, could track a suspect through a single large space but also track him or her later on. If a suspect is spotted by cameras in a mall and then hours later at a nearby grocery store, the AI could potentially draw from this information to locate the suspect in areas between those two points as well, giving law enforcement a more complete picture of the suspect's movements.

The Surveillance Issue

The ACLU is less concerned with the “how” of Rekognition and more with the “who” and “what.” It filed a public records request to learn more about Rekognition and found what Cagle and Ozer's blog points to as troubling evidence of Amazon colluding with government agencies to push toward a "surveillance state." According to records obtained by the ACLU, Amazon readily offers support to government agencies deploying Rekognition, including providing troubleshooting and best practices. It also solicits feedback from law enforcement on how it may improve Rekognition's capabilities for the use case.

The records obtained by the ACLU show that several “fusion centers” (collaborative efforts among two or more agencies that share resources, expertise, and/or information related to criminal activity and investigations)—as well as law enforcement agencies in California and Arizona—have reached out to the Washington County Sheriff's Office regarding Rekognition.

Comments (4)

Please log in or to post comments.
  • Oldest First
  • Newest First
Loading Comments...