Insufficient security in ML deployment can create catastrophic vulnerabilities in the AI process. It’s time to get ahead of it.
Stained Glass transforms sensitive data at every Ml touchpoint. Use Protopia AI to securely deploy AI models in the inferencing stage, which is often unprotected
STG for ML Inferencing Solve for data privacy in the ML deployment stage for any data type
Companies lack consistent protection for AI deployment. Stained glass bridges the gap with transformations on any data-type
Choose your data type
Watch Protopia AI in action with the face recognition task below.
Today, the face recognition model would see the data record in its raw form as it appears to the left. This exposes all the information in plain/raw form to the prediction task. With Protopia AI’s Stained Glass Transfrom™ in place, the model will see a completely transformed data record as is shown on the right, but can still make accurate predictions with much less information.
Watch Protopia help an ML task make a prediction on a loan application with the example below.
Traditionally, a model would look at every piece of data in the data record and a combination of all those pieces of information in raw form. This approach enables the leaking of sensitive information about the application such as PII. With Protopia in place, the model will read a completely transformed and randomized piece of data but still be able to predict the outcome with the same level of accuracy.
Notice how the personally sensitive attributes on the left are obfuscated on the right with the stochastic embedded representation. In addition, the reconstructed information of text and numbers does not match the actual data depicted on the left. That’s a transformation that automatically occurs in this process.
Watch Protopia in action with an NLP task performing phishing email detection.
However, current NLP tasks have the opportunity to consume data that contains too much sensitive information such as email content. The left side below shows what the tokenized version of emails look like today when using NLP services. On the right, using Protopia’s stochastic transformations, you’ll see how we help protect sensitive information while still enabling an NLP task to identify spam and work accurately.
- Visual Data
-
Watch Protopia AI in action with the face recognition task below.
Today, the face recognition model would see the data record in its raw form as it appears to the left. This exposes all the information in plain/raw form to the prediction task. With Protopia AI’s Stained Glass Transfrom™ in place, the model will see a completely transformed data record as is shown on the right, but can still make accurate predictions with much less information.
- Structured/Tabular Data
-
Watch Protopia help an ML task make a prediction on a loan application with the example below.
Traditionally, a model would look at every piece of data in the data record and a combination of all those pieces of information in raw form. This approach enables the leaking of sensitive information about the application such as PII. With Protopia in place, the model will read a completely transformed and randomized piece of data but still be able to predict the outcome with the same level of accuracy.
Notice how the personally sensitive attributes on the left are obfuscated on the right with the stochastic embedded representation. In addition, the reconstructed information of text and numbers does not match the actual data depicted on the left. That’s a transformation that automatically occurs in this process.
- Natural Language Processing
-
Watch Protopia in action with an NLP task performing phishing email detection.
However, current NLP tasks have the opportunity to consume data that contains too much sensitive information such as email content. The left side below shows what the tokenized version of emails look like today when using NLP services. On the right, using Protopia’s stochastic transformations, you’ll see how we help protect sensitive information while still enabling an NLP task to identify spam and work accurately.
US Navy: AI Facial Recognition
The US Navy used AI facial recognition at base checkpoints, which posed a risk to sensitive images of base personnel, structures, and equipment.
Protopia AI partnered with NetApp and implemented Stained Glass Transform™ for real-time face recognition during the Trident Warrior exercise. The solution transformed faces into Randomized Re-Representations only readable by the AI model, using maximal curated noise with near-perfect accuracy to only identify persons of interest and block out sensitive items in the background.
