AI-Assisted Audio/Video Redaction
Description
This use case intends to solve the problem of the labor-intensive process of redacting audio and video evidence.
Detailed example
The AI output in this use case are redactions to the media file. Users will further edit the redacted media prior to exporting the final redacted file to ensure completeness. Homeland Security Investigations conducts a human review of each frame within redacted files prior to distribution.
AI / analytics pattern
Computer Vision: AI that processes and interprets visual data (e.g., images and videos).
Automation level / stage
c) Deployed – The use case is being actively authorized or utilized to support the functions or mission of an agency.
Expected benefit
The AI-Assisted Audio/Video Redaction tool is used to reduce the manual effort required to redact audio and video evidence used during an investigation and subsequent legal proceedings.
Audit / financial statement impact
The AI-Assisted Audio/Video Redaction use case does not meet the definition of a High-Impact category under OMB M-25-21 due to its narrowly defined scope, reliance on human oversight, and non-decision-making role. This tool is designed to assist Homeland Security Investigations (HSI) by partially automating the redaction of audio and video evidence, such as detecting and obscuring faces, objects, or sensitive information (e.g., license plates, PII), to protect individuals’ identities during investigations and legal proceedings. Importantly, the tool does not perform 1:1 facial matching or identification, and its outputs are strictly limited to redactions, which are subject to comprehensive human review and editing before finalization. This ensures that the AI’s role is supportive rather than determinative, with no direct impact on investigative decisions or legal outcomes.
Controls / human review
ATO: No; PIA: https://www.dhs.gov/sites/default/files/2024-03/24_0307_priv_pia-ice-066a-pia-update.pdf
Data needed
All training sets used for the model are from public and private collections of images. AI models are trained using a combination of real-world and synthetic datasets collected from publicly available sources. These datasets are curated to represent a broad range of conditions, including edge cases such as occlusions and poor lighting, to improve detection accuracy across varied scenarios.