Facial Recognition Technology

The Arvada Police Department (APD) has in the past successfully used facial recognition technology as an investigative tool. This technology can be used to help law enforcement generate leads as to the identity of unknown criminal suspects or missing or otherwise at-risk persons. Facial recognition technology does not replace other investigative techniques. Officers may not use facial recognition technology as the sole basis to establish probable cause in a criminal investigation.

In 2022, the State legislature passed Senate Bill 22-113, which added sections 24-18-301 through 309 to the Colorado Revised Statutes. These laws govern the use of facial recognition technology by governmental agencies, including law enforcement. Governmental agencies are now required to take several steps before utilizing facial recognition technology. Among these steps are requirements that the agency post a draft “accountability report” to its website. This accountability report documents various information about the specific facial recognition technology application that the agency will utilize, and how it will be used. The APD is working to re-implement use of facial recognition technology in accordance with statutory requirements. Therefore, the APD posts its draft accountability report below.

The APD is seeking public comment on its accountability report. Further, the APD invites interested community members to attend any of three community meetings regarding the APD’s use of facial recognition technology to be held at APD community stations as follows:

  • September 25, 2023, 6pm to 7pm at Arvada Police Headquarters, Anne Campbell Room, 8101 Ralston Road
  • September 26, 2023, 6pm to 7pm at Whisper Creek Community Station, 14360 West 89th Drive
  • September 27, 2023, 6pm to 7pm at Lake Arbor Community Station, 8110 Vance Drive

You may also contact us at 720-898-6668 or email at [email protected] to provide comment.

Draft Facial Recognition Accountability Report

Arvada Police Department

Facial Recognition Accountability Report

LexisNexis Lumen Facial Recognition Service 

The Arvada Police Department (APD) has prepared this Accountability Report pursuant to the requirements of Colorado Senate Bill 22-113, as codified in Colorado Revised Statutes (CRS) §§ 24-18-301 through 309.   This Accountability Report discusses information required by CRS §§ 24-18-302(2)(a-h).  Statutory references are noted where applicable.    

The APD will activate the facial recognition functionality (hereafter referred to as the facial recognition service or “FRS”), facilitated by Rank One Computing Corporation, within the LexisNexis Lumen software platform.  Lumen is an investigative application that utilizes the criminal justice information shared between the law enforcement member agencies of the Colorado Information Sharing Consortium (CISC).  The FRS within the Lumen platform is provided by Denver-based, Rank One Computing Corporation’s (ROC) SDK version 2.2.1 algorithm software (CRS § 24-18-302(2)(a)(I)).    

This software uses state-of-the-art facial recognition technology to match the face in a user-uploaded image to mugshot images from CISC member agency records.  It is designed to be used in ways that ultimately reduce crime, fraud, and to enhance community safety. All APD use of FRS shall be for official law enforcement purposes only and considered law enforcement sensitive information (CRS § 24-18-302(2)(a)(II)). 

 In accordance with CRS § 24-18-307, the APD will use FRS information only as investigative leads and will utilize any FRS results only in conjunction with other leads and evidence.

 Technical Description & Intended Use

Lumen FRS may be used in an investigation to help identify potential suspects by comparing a single user-uploaded image (the “probe image”) of an unknown suspect to a collection of gallery facial images provided by the CISC.  Lumen FRS provides multiple results, each with a given match score generated by the ROC SDK’s facial recognition algorithms.  The match score is designed to indicate the likelihood of the probe image matching a given result (CRS § 24-18-302(2)(b)(I)).

Capabilities & Function

The Lumen FRS helps solve crimes after-the-fact by matching photos obtained by a government customer, of suspects, persons of interest to a law enforcement investigation, and victims or possible victims of crimes against images of known persons contained within the CISC member-agencies records. Specifically, the Lumen FRS uses a machine-learning facial recognition algorithm to initiate a search between the face in probe image against the images contained within the criminal justice records available only to members of the CISC.  At that point, law enforcement personnel make independent assessments to determine if there is a match between the probe image and images scored high within Lumen. Each decision about a possible match is made by an FRS-trained member of the APD (CRS § 24-18-302(2)(d)(I)).

The core facial recognition algorithms depend primarily on the image quality of the probe image and gallery images and on the robustness of the algorithm development process. The primary factors of image quality are capture conditions, including camera sensor quality, field of focus, glare, blur, low light, high contrast, variable lighting, height of the camera, pose of the subject and occlusions between the camera and the subject face. Algorithms are developed by processing training data through machine learning architectures and iteratively testing accuracy on data that represents real-world conditions. Accuracy of a match score may be impacted by poor image quality of the probe image and/or gallery image or to the extent that operational data is fundamentally dissimilar to training data and/or testing data selected in the research and development process (CRS § 24-18-302(2)(a)(II)).  

Decision Making

Lumen FRS is intended only to support APD investigations.  It does not make any decisions as to whether the probe image is a match to a database image.  Each decision about identification is made by a member of the APD, not by an automated process.  APD personnel must review the results for every search to enable human review and independent verification.  Probable cause determinations may not be based solely on these identifications. In addition, all results will be peer reviewed by another FRS-trained APD member prior to utilizing any information obtained (CRS § 24-18-302(2)(a)(I)).

Intended Use and Benefits

Lumen FRS is intended to enhance the APD’s investigative capabilities.  This type of facial recognition technology automates the process necessary to locate potential matches between a probe image and thousands of criminal justice record images that would otherwise require a manual search by a human.  The facial recognition algorithm will rank potential matches in a manner that allows for a simplified process of human review (CRS § 24-18-302(2)(a)(I)).

When provided a probe image to search against a collection of gallery images, Lumen FRS returns multiple results, sorted by the highest match score generated by the ROC SDK’s facial recognition algorithms, Once Lumen FRS provides a list of results, a human investigator must review the results before making any determination of a possible match. A possible match determination may be used as an investigative lead that is treated in a similar manner as an anonymous tip. In particular, the investigative lead does not supply adequate probable cause to make an arrest without additional evidence (CRS § 24-18-302(2)(a)(III)). 

The intended benefit of using the Lumen FRS is to generate investigative leads for further investigation with the hope of solving unsolved crimes. By way of example, in comparable use by the New York City Police Department (NYPD) since 2011, the NYPD has successfully used facial recognition to identify suspects whose images have been captured by cameras at robberies, burglaries, assaults, shootings, and other crimes. In 2019, the NYPD Facial Identification Section received 9,850 requests for comparison and identified 2,510 possible matches, with no known instance which a person was falsely arrested on the basis of a facial recognition match (CRS § 24-18-302(2)(a)(I)).[1]

The use of FRS will assist the Arvada Police Department in its efforts to:

  • Increase public safety;
  • Minimize the threat and risk of injury to specific individuals;
  • Minimize the threat and risk of physical injury or liability to law enforcement and others responsible for public protection, safety, and health;
  • Minimize the potential risks to individual privacy and civil rights, civil liberties, and other legally protected interests;
  • Reduce risks of bias and prejudice impacting criminal justice processes;
  • Protect the integrity of criminal investigations and justice system processes and information; and
  • Effectively use public resources allocated to police services (CRS § 24-18-302(2)(a)(I)).

Data Inputs and Generation 

The Lumen FRS uses the following types of data inputs:

  • User submitted probe images and associated information identifying the purpose of the search (such as case number and type of crime).
  • The gallery facial image data is collected by the CISC from its member agencies, the national NCIS Law Enforcement Information Exchange (LInX) and the FBI’s N-DEx national information sharing system.

The Lumen FRS generates a template of each facial image, which is a mathematical model of the unique subject which may be compared to templates generated from other images to produce a match score. For each facial image, the tool also generates metadata including pitch, yaw, image quality estimations and facial analytics like age, gender, geographic origin, emotion, facial hair, glasses and mask estimations (CRS § 24-18-302(2)(b)(II)).

Data Management, Training and Use Policy 

The APD will follow statutory requirements described in CRS §§ 24-18-301 through 309, in conjunction with APD Policy 703 (CRS § 24-18-302(2)(d)(VIII)).  The APD will follow the below guidelines regarding data management, training and the authorized use of the facial recognition service.

 Data Minimization

The features and function of the Lumen FRS effectively reduces the risk of inadvertent access to data by APD personnel. Lumen FRS searches only criminal justice records available to CJIS-certified law enforcement personnel of CISC member agencies.  The criminal justice records available in the FRS are subject to the retention policies of the owner agencies (CRS § 24-18-302(2)(d)(III)).

 Data Integrity and Retention 

The Police Operations Manager or their designee will oversee Lumen FRS permissions for the APD.  The Police Operations Manager or their designee will have the capability to audit and review any and all usage of FRS by any authorized member of the department.   The audit will include all user’s activity, such as user log ins and log outs, each user’s activity in detail, what commands were issued to the system, and what records or files were accessed (CRS § 24-18-302(2)(d)(IV)). 

All information obtained from the Lumen FRS by a member of the APD will be documented in the related investigative case report and retained in accordance with APD records management policies.    

LexisNexis/Lumen employees will not have access to and will not review APD search history within the Lumen FRS without the express permission of APD, or as required by law or court order. This ensures that FRS investigative data will remain confidential. APD FRS data will not be shared with non-law enforcement agencies (CRS § 24-18-302(2)(d)(VI)(B)). 

All information available within the Lumen investigative platform, including the FRS, is purged according to the retention schedule and policies set by the owner agency.  Thus, information made available to other CISC member agencies by the APD is purged from the Lumen investigative platform when its retention period expires in APD’s record management system.  APD FRS data will be retained in accordance with the retention schedule applicable to the underlying criminal offense in accordance with the Municipal Records Retention Schedule adopted by the City of Arvada, and the evidence preservation requirements applicable to each case (CRS § 24-18-302(2)(d)(IV)).

 Usage Rules and Requirements

Access to FRS search results will be provided only to APD members who are authorized to have access and have completed applicable training. Authorized access to the APD facial recognition software will be granted only to personnel whose positions and job duties require such access for investigative purposes.  The Police Operations Manager or designee shall grant and audit all user access, following the required account approval.  All facial recognition users shall be required to have individual access for use of the FRS (CRS § 24-18-302(2)(d)(II)).  

Approved FRS operators will analyze, review, and evaluate the quality and suitability of probe images, to include factors such as the angle of the face image, level of detail, illumination, size of the face image, and other factors affecting a probe image prior to performing an FRS search (CRS § 24-18-302(2)(d)(I)).   

Original probe images shall not be altered, changed, or modified to protect the integrity of the image. Any enhancements made to a probe image will be made on copies, saved as a separate image, and documentation will indicate what enhancements were made, including the date and time of change.  The resulting images, if any, shall be manually compared with the probe image by the person conducting the comparison.

Any upload of a probe image, query, or request shall include the name of the agency/requestor, name of the person completing the request, date and time the request was completed, case number and reason for the request. This information will be logged, tracked and available for auditing and review in the related case report. 

The APD and all authorized facial recognition users shall comply with all requirements stipulated in any APD agreements related to authorized facial recognition enrollment databases (CRS § 24-18-302(2)(d)(I)).

Images accessed by the APD for FRS searches are not maintained or owned by the APD and are subject to the retention policies of the respective enrollment databases authorized to maintain those images.  

In accordance with CRS § 24-18-303 and APD Policy 703, members shall disclose the use of FRS to prosecutors for disclosure to a criminal defendant in a timely manner prior to trial.

Data Security 

Facial recognition data is stored securely on Lumen servers, and access is limited to authorized users within Lumen. FRS data will not be shared with non-law enforcement agencies (CRS § 24-18-302(2)(d)(VI)(B)).
Lumen is a web-based software and not an application which needs to be downloaded to any City of Arvada computers.   Any records exported by APD members shall be immediately uploaded to the department’s CJIS-compliant records and/or evidence management system (CRS § 24-18-302(2)(d)(VI)(A)).

 Training Procedures 

The APD will provide training to all authorized FRS users. This training will be arranged and documented by the Police Operations Manager.  FRS account access will not be created or provided until training has been completed.

Training will include the use of FRS technology as well as a specific review and acknowledgment of all elements of APD policy and this Accountability Report.

 In accordance with CRS § 24-18-305, FRS training will at a minimum include:

  1. The capabilities and limitations of the facial recognition service.
  2. Procedures to interpret and act on the output of the facial recognition service; and
  3. To the extent applicable to the deployment context, the meaningful human review requirement for decisions that produce legal effects concerning individuals or similarly significant effects concerning individuals.

The use of each authorized FRS enrollment database will include specific training that includes the following:

  1. each authorized user will access only their individual account;
  2. the authorized user shall document in a case report all required information to support the authorized use of facial recognition satisfying an official law enforcement purpose;
  3. how a lawfully obtained probe image of a subject meeting the required authorized use is uploaded to the system;
  4. the software automatically compares the probe image to gallery images within the repository;
  5. results of the comparison are returned and provide a potential investigative lead.

Updated training shall be identified with any policy revisions or updates to the FRS (CRS § 24-18-302(2)(d)(VII)).

Testing Procedures 

As required by CRS § 24-18-304(4), Rank One Computing has previously submitted the ROC SDK for testing in the following series of the National Institute of Standards and Technology (NIST) Face Recognition Vendor Test (FRVT) (CRS § 24-18-302(2)(e)):

1:1 Verification -                https://pages.nist.gov/frvt/html/frvt11.html

1: N Identification -            https://pages.nist.gov/frvt/html/frvt1N.html

Quality Assessment -        https://pages.nist.gov/frvt/html/frvt_quality.html

Demographic Effects -       https://pages.nist.gov/frvt/html/frvt_demographics.html

Paperless Travel -             https://pages.nist.gov/frvt/html/frvt_paperless_travel.html

Presentation Attack Detection -    https://pages.nist.gov/frvt/html/frvt_pad.html

Accuracy and Impact 

Test Results

Rank One Computing’s SDK facial recognition algorithm was submitted to the National Institute of Standardization and Technology (NIST) Face Recognition Vendor Test (FRVT) for 1:1 Verification. In that test, ROC’s SDK facial recognition algorithm ranked No. 10 in the world out of 478 total entries and was the top ranked entry from the United States (CRS § 24-18-302(2)(f)).[2]

Bias and Inaccuracy

In the NIST Demographic Effects series[3] the ROC SDK ranked 8th worldwide across all 70 sub-populations of the NIST test data, with the lowest scoring demographic being West African females aged 65-99 years old (0.01871% false match rate).  

Civil Rights Impact

The potential impact of a false match, including on protected subpopulations, is mitigated by the human investigator review requirement as well as by the requirement to develop additional evidence prior to making an arrest.  The direct impact of an erroneously high match score from the ROC SDK is that a candidate would rank higher on the list of results returned by Lumen FRS for human investigator review.  The human investigator would then apply their skills, training, and experience in the facial examination to closely review the unique characteristics of each of the candidates on the list.  The human investigator may select one of the candidates from the list of results and make a possible match determination based on the similarity of facial characteristics between the candidate and suspect image, or instead may determine that none of the candidates from the list of results are a possible match (CRS § 24-18-302(2)(g)).  

If the false match eludes both the ROC SDK and the human investigator, it could become an investigative lead, triggering additional investigation into the relevant candidate.  In the absence of additional evidence, erroneous investigative leads do not result in a false arrest.  By way of example, statistics provided by the NYPD show that the agency uses facial recognition tens of thousands of times each year without a known instance of false arrest. [4]

APD personnel with access to the Lumen FRS are required to document the related case number and crime type prior to initiating a search, affirmatively representing that the search is conducted for the purpose of investigating a crime that has been committed. As a result, the use of the Lumen FRS for “fishing expeditions” or monitoring persons engaged in lawful activities is curtailed.  This information will be verified by mandatory supervisory review of cases which include user inputs into the facial recognition service to ensure user compliance (CRS § 24-18-302(2)(d)(III)).

APD usage of the Lumen FRS is unlikely to have a negative impact on the civil rights, liberties, privacy, or on marginalized communities of the people of the State of Colorado. ROC’s SDK algorithm achieved greater than 99% accuracy across all demographic groups on NIST’s FRVT Demographic Effects in Face Recognition test program; thus, disparate impact on marginalized communities is likely to be negligible. The APD will not utilize Lumen FRS if at any time independent testing determines its rate of error or false matches exceeds one percent (CRS § 24-18-302(2)(f)).

The APD has clear policies set forth in APD Policy Manual 402 and 703 governing the use of FRS and prohibiting investigations into individuals based in whole or in part on a person’s actual or perceived race, ethnicity, gender, national origin, language preference, religion, sexual orientation, gender identity, age or disability, unless that investigation is based on a reliable suspect-specific description of the individual that includes other non-demographic identifying characteristics (CRS § 24-18-302(2)(d)(VIII).

Public Feedback

The Arvada City Council, by resolution dated March 6, 2023, approved the APD’s intent to implement and utilize FRS. 

As required by CRS § 24-18-302(3)(a-c), the APD will hold public meetings to receive comments and feedback regarding the APD’s use of FRS September 18-21, 2023, at APD stations.  

The APD will consider the public comments and feedback provided at these meetings. In addition, the APD will receive and respond to community feedback on its use of FRS through other methods of communication, including but not limited to, US Mail, the APD citizen comment line, or any other means of in-person or electronic communications (CRS § 24-18-302(2)(h)).  

The APD will monitor its utilization of FRS in investigative processes through various supervisor reviews and case analysis. Complaints or commendations regarding the APD's use of FRS will be routed to the Police Operations Manager or the Internal Affairs Unit as necessary under the circumstances (CRS § 24-18-302(2)(g)).

[1] https://www.nyc.gov/site/nypd/about/about-nypd/equipment-tech/facial-recognition.page

[2] Please see full report at https://pages.nist.gov/frvt/html/frvt11.html