Ben Ha is the Solutions Architect Director for Veritone’s Government, Legal and Compliance division. Ben has over 15 years of experience in the software industry, serving primarily in a technical pre-sales role. Ben has been working with clients in the government and legal space for the last 4 years.
Veritone designs human-centered AI solutions. Veritone’s software and services empower individuals at many of the world’s largest and most recognizable brands to run more efficiently, accelerate decision making and increase profitability.
How does Veritone’s iDEMS integrate with existing law enforcement systems, and what specific efficiencies does it introduce?
Law enforcement agencies’ (LEAs) existing systems typically have data from many different sources, like body-worn camera systems, video management systems and other cameras and devices. iDEMS allows LEAs to build connections in those existing systems with an API or other integration pathways. It then virtualizes over the top of those systems, permitting law enforcement to keep the master data where it is in the source systems. Inside the Veritone Investigate application, the user has access to a low-resolution proxy file they can leverage for viewing, sharing, searching, analyzing, etc. Because the data is in one central location, it is easier for the user to go through the investigative process without switching between siloed applications.
Veritone Investigate also allows the user to leverage AI cognition to analyze what is inside the content itself. In other words, LEAs can use AI to structure unstructured data, providing metadata information that makes finding things much easier. Most systems simply act as data storage and do not contain information about the words spoken or the faces or objects inside the content. With Investigate and the iDEMS solution, AI is natively built-in and runs automatically upon ingestion, eliminating the need to manually watch or listen to content to obtain context, accelerating the investigative process.
What are the technical requirements for law enforcement agencies to implement Veritone’s iDEMS?
LEAs don’t need to possess significant technical requirements to implement Veritone’s iDEMS – in fact, the solution will work with almost any sized LEA regardless of what systems they do or do not have in place. Because Veritone has ingestion adapters that can connect with various APIs, the only thing the LEA will need is someone with access to those existing systems. Also, iDEMS is cloud-based, and the LEA will need a high-speed internet connection and a modern web browser.
Can you provide more details on how Veritone Track differentiates from traditional facial recognition technologies in terms of accuracy and efficiency?
Traditional facial recognition relies on visible facial features (eyes, nose, mouth, etc.) to identify a person of interest. The issue is that if the video does not capture the person’s face, the technology cannot identify or track that individual. For example, if the footage only captures someone’s back, the person’s face is covered by a mask or hoodie, or the video doesn’t have an optimal angle of the face, the facial recognition won’t work.
Alternatively, Veritone Track treats potential persons of interest as objects in a process known as human-like objects (HLOs). Through HLOs, Veritone Track can build a unique “person print” of that individual based on visually distinguishing attributes. These visually distinguishable attributes could be a hat, glasses, backpack or if they are carrying something in their hand, even the color contrast between their clothing and shoes. It also considers the person’s body type, e.g., arm length, stature, weight, etc.
After building that person print, Veritone Track incorporates good old-fashioned police work through a human-in-the-loop that reviews and verifies potential matches. Ultimately, this method is more accurate and efficient than traditional facial recognition technologies.
How does the use of human-like objects (HLOs) in Veritone Track enhance the identification process compared to using facial recognition?
Leveraging HLOs enhances the identification process because it doesn’t require the LEA to have access to the same variables as traditional facial recognition, i.e., a fully visible human face. Veritone Track is flexible in that it will use whatever information is available regardless of the quality of the footage, the resolution or the angle (high up on the ceiling or at eye level) of the camera. Despite the advantages of Veritone Track, it and facial recognition are not mutually exclusive – LEAs can use both technologies simultaneously. For example, LEAs could use Veritone Track to construct a person print from large amounts of lower-quality video while running facial recognition on video samples of front-facing shots of a potential person of interest.
How does Veritone’s AI-powered system help in speeding up investigations while maintaining high standards of evidence handling?
Veritone Investigate, Veritone Track, or all of Veritone’s public sector applications use AI to dramatically accelerate manual processes for LEAs, reducing weeks or days’ worth of work into a few hours, which is increasingly critical amid ongoing staffing shortages. Despite this accelerated speed, Veritone maintains high standards of evidence handling by not totally trusting AI outputs. These solutions leave the final say to the human investigator to review the final results. Veritone’s technology also enables humans to conform to high standards of evidence handling and chain of custody. Likewise, they have built-in audit trails, so the LEA can see how the investigator arrived at the final result. Put simply, AI does not replace humans – it merely enhances their capabilities.
AI in law enforcement raises concerns about wrongful persecution of minorities, especially with cities like Detroit, Michigan experiencing multiple wrongful arrests in less than 1 year. How does Veritone address these ethical challenges?
First, Veritone always uses guardrails and safety measures to minimize the possibility of wrongful persecution. For instance, Veritone Track does not use biometric markers such as facial features to build person prints but relies on clothing, body type, etc. Second, these tools never scrape the internet, social media or huge databases like a Passport Agency to obtain data. When an LEA uses our solutions in an active case or investigation, it can only compare uploaded photo or video evidence against a database of known offenders with arrest records. In the case of what happened in Detroit, Michigan, law enforcement used a solution that grabbed data from across the internet without a human investigator being “in the loop” to validate the results, resulting in wrongful persecution of innocent citizens.
Can you elaborate on how Veritone’s AI ensures the accuracy of the leads generated?
Veritone’s AI generates potential leads that human investigators can pursue. While the AI provides the investigator with helpful findings and results, the person still makes the final decision. Again, the Detroit, Michigan, case saw law enforcement trusting facial recognition alone to do the job. This blind trust was ultimately problematic as these models relied on data that resulted in demographically or racially associated biases.
Moreover, the data Veritone chooses to train its AI engines and models are representative of the content. Before training the data, Veritone will redact sensitive video and audio elements from sources like body-worn cameras, in-car video, CCTV footage, etc., or use publicly available non-sensitive data. Likewise, Veritone will validate results with customer feedback for continuous improvement.
How does Veritone handle the potential for AI to perpetuate existing biases within law enforcement data?
Veritone uses a multiple-model approach that works with many different third-party providers to obtain a larger perspective rather than relying purely on one AI model. In particular, this method allows Veritone to standardize within a given category of AI cognition, such as transcription, translation, facial recognition, object detection or text recognition. By leveraging the “wisdom of the crowd,” Veritone can run the same content against multiple models within the same category of AI cognition to help guard against biases.
What steps are taken to ensure that Veritone’s AI applications do not infringe on privacy rights?
There are two best practices Veritone’s AI applications follow to ensure they do not infringe on privacy rights. One: the customer’s data remains the customer’s data at all times. They have the right to manage, delete or do whatever they want with their data. Although the customer’s data runs in Veritone’s secure cloud-hosted environment, they retain full ownership. Two: Veritone never uses the customer’s data without their permission or consent. In particular, Veritone does not use the customer’s data to retrain AI models. Security and privacy are of the utmost importance, and customers will only ever work with pre-trained models that use data redacted of all of its sensitive, biometric and personally identifiable information.
How does Veritone balance the need for rapid technological advancement with ethical considerations and societal impact?
When developing AI at a rapid pace, the tendency is to use as much data as possible and continually harvest it to improve and grow. While such an approach does tend to result in accelerated maturity of the AI model, it opens up various ethical, privacy and societal concerns.
To that end, Veritone is always looking for the best-of-breed AI. During the generative AI craze, Veritone had early access to technology from OpenAI and other partners. However, instead of pushing ahead and deploying new solutions immediately, we asked, “How will our customers actually use AI within a proper use case?” In other words, after examining the mission and pain points of LEAs, we determined how to apply Generative AI in a responsible way that kept humans at the center while allowing users to achieve their goals and overcome challenges.
For example, Veritone Investigate features a private and network-isolated large language model that can summarize spoken conversations or content. If a body-worn camera captures an incident or an investigator interviews someone, Veritone Investigate can transcribe that content and automatically summarize it, which is very helpful for detectives or investigators who need to provide a summary of an entire interview in a short paragraph to the DA or prosecution. Nevertheless, the person still has the chance to review the AI-generated output to make necessary edits and changes before submission.
Thank you for the great interview, readers who wish to learn more should visit Veritone.
Credit: Source link
