AI/ML systems security
AI/ML is essential for modern businesses. It includes analysing customer behaviour and using data processing methods to deliver new value to customers. Novel technologies bring novel threats or exacerbate existing ones, and since they’re data-related and math-heavy, we’re often asked by our customers to help tackle them.
We design and build cryptographic solutions for AI/ML-driven businesses, including ML model security, DRM-like schemes and protecting ML models against reverse engineering attacks.
Typical challenges for AI/ML security
Hard to balance
Data engineers want more real-world data to train their models, to run analytics. While security team wants to use minimum production data.
Protecting ML models
When a unique competitive advantage is condensed in the ML model or any sophisticated data processing contraption, it becomes a sensitive security asset which requires protection.
Customer privacy rights
ML-driven products often deal with sensitive consumer data and trigger privacy concerns that require tweaking ML architecture to get it right.
Data security concerns
ML needing to operate on sensitive data under complex security constraints requires building security inside ML infrastructure and around it.
Modern solutions
ML model security
DRM security schemes for ML models to run only on authorized applications without risks of leakage.
Differential privacy
Applying various privacy models (differential privacy, K-anonymity) to practical goals requires careful planning, implementation and validation.
Inference/statistical security
Understanding statistical security consequences and other indirect security risks and building data processing that mitigates those risks.
Our offerings
// Relevant products
Themis
A CROSS PLATFORM CRYPTO LIBRARY
To be announced
// Custom design and implementation
DRM-like ML-models protection
Differential privacy systems
Security layers for complex use cases
// Consulting
Protection against statistical attacks
Pragmatic security
Product security strategy
Have a question? Get a human to answer it!
How we make a difference
Vast experience and expertise
As a cryptographic R&D team, we've built custom data security controls in different contexts, and have vast expertise in understanding practical threat models and failure scenarios, and designing sound security controls against them.
Relevant experience in ML
We’ve done security for award-winning ML products that are used by tens of millions of users, sophisticated industrial data-mining solutions and planet-wide services.
Built to last
Our approach is based on backwards compatibility, maintainability and support. You never lose encrypted data because some npm package became outdated.
For innovators, by innovators
We've started Cossack Labs to develop new tools and methods for protecting the data and enabling novel solutions to emerging problems — so that at the edge of your innovation, you’ve already got fitting tools handy.
Contact us
There are many ways we can help: with our products, bespoke solutions, and engineering services. Leave your contact information to connect with our team:
Interesting blogposts
Security of React Native libraries: the bad, the worse and the ugly
How to select a secure React Native library for your app. Sort out improper platform usage, easy to misuse API, deprecated and abandoned libraries – check our research of the React Native ecosystem security.
Implementing End-to-End encryption in Bear App
Helping Bear app implement note encryption for their vast existing user base. Balancing usability, security, and mobile platforms' restrictions.
Audit logs security: cryptographically signed tamper-proof logs
Why crypto signed audit logs are essential for security software and how we’ve built-in secure audit logging in Acra for defense in-depth.