Veritone Redact is used by law enforcement agencies and legal teams to quickly obfuscate sensitive information from photographic and video-based evidence. The most common use of the product is to redact faces of identifiable individuals.
To do so, Veritone developed a proprietary neural network-based model trained to recognize human heads within a scene. To be clear, Redact does not employ any facial recognition technology, nor a means to identify personally identifiable attributes such as skin or hair color, race, or gender. Instead, the product is designed to find human heads within a scene and provide the operator with an efficient means to determine which heads should be redacted from the resulting output.
The neural network-based model was architected in a two-step process:
- First, expose the neural net to millions of sample images with diverse imagery
- Next, introduce a secondary training dataset of images which includes human heads
The secondary training set was introduced to inform the neural net on the type of imagery (human heads) the model should be able to subsequently identify. Given those two sets of data, the neural net figures out what the identifying characteristics of a head are and trains itself to accurately find them. The result is a model that is capable of detecting heads based on an object’s overall shape, relation to other objects in the scene and not limited to the inspection of attributes specific to a region on the head itself (face) such as eyes, lips or skin tone.