FAQ’s

Classifier Capabilities

Yes, the VigilAI CAID Classifier works on both video and still images.
Application of the Classifier to video streaming is on our product roadmap.
An initial pro-bono trial was carried out in 2016, with UK Home Office funding since 2017 deployed to UK police and national law enforcement organizations in 2018.

Use Of Training Data

We have a permissively sourced set of legal adult content from a very large vendor, as well as a hand-curated benign dataset that has been built over the years.

In addition to the CAID Training Data, there is:- Training data categorized by CSA Investigators.
– Fine Grained Categorization – not just detection.
– Extensive performance assessment (larger dataset means better training and better evaluation).
– The fact it is built by a team that includes computer vision experts and CSA Investigation expertise (we know the problem and the solution).
– Video classifier to go along with the image classifier.

It falls in the many millions.

We were approved and selected by the Home Office for deployment when technically possible across UK law enforcement and various national agencies.

Yes – we have an outgoing API, pushing out classification with confidence scores that can inform your custom workflow trafficking. We also have front end options.

Technical Details

The Classifier outputs are in the form of confidence scores, which with the right governance structure is ideal for decision making.

No – the classifier is not trained on your data and no CSAM or even performance data is shared back to Krunam. Instead, we prefer to have regular performance check-ins where you share information that you are comfortable sharing with Krunam so that we can increase performance.

Yes, but you may need to spin it up on more machines based on scale. It is deployed via Docker, allowing for easy scaling.

Testing on your hardware with your data is part of our typical engagement path – please contact us if you would like to test the VigilAI CAID Classifier.

– The “developer” version doesn’t have any visual display; it just provides confidence scores for each image which are then integrated into your existing workflow and visualization system.

– We recommend using the technology to rank images for moderation, for example on a class by class basis.

– Front-end options are available.

This is a giant leap forward from perceptual hashing – moving from only known images to finding the entire class of CSAM through computer vision. For the developing AI solutions, that’s where our dataset, training procedure and CV/AI expertise set us apart.

Our classifier uses a computer vision algorithm that was trained using state of the art AI and deep learning techniques. The technology identifies “hallmarks” of CSAM from being trained on the CAID dataset, the pre-eminent collection of privacy safe, legally obtained CSAM.

Various deployment options are available:

– We have Windows and Linux versions

– We can deploy via Docker for scalability

– We have outbound API’s to build into your systems. If you don’t have content moderating and governance systems we have options for you as well.

Yes, fully. We never actually house the CSAM data – and all of our non-CSAM data is legally obtained. We have no individual identification ability built into the system, as it looks at classes of behavior and conduct.

Once you are under an NDA, we can share our internal testing, which is rigorous and multi-layered. We have held back a sizable portion of our CSAM dataset for testing, and built a robust proprietary hard negative dataset.

Performance is high enough that our law enforcement partners are using it in real operational work to speed up detection and categorization, and they also don’t want knowledge of the specific results out in the public domain. Please contact us to discuss this further.