Refortif offers a CLI tool that obfuscates AI model weight files into an unreadable format and a secured runtime that executes these transformed models without ever exposing the original weights in memory or on disk. This protects proprietary models from GPU memory dumping and file extraction attacks while maintaining native inference speed, enabling AI developers and enterprises to distribute models safely.
Funding
Funding not disclosed
Founders
Product
Problem
AI model owners risk losing proprietary weight files when models are distributed, as attackers can dump GPU memory or extract model files, leading to permanent loss of intellectual property and enabling unlicensed use.
Solution
Refortif provides a protection pipeline that mangles model weights into an unreadable form before distribution. A secured runtime, embedded with a private key, executes the obfuscated model without ever materializing the original weights on disk or in RAM. The process is performed via a CLI tool, after which only the transformed weights are shipped. Because the runtime operates natively, inference speed remains comparable to unprotected models, allowing seamless integration into existing deployment workflows.
Target Audience
Primary customers are AI developers, enterprises, and SaaS providers that distribute proprietary models and need to safeguard their intellectual property against unauthorized access.
Features
- CLI-based mangler that converts original weights into an obfuscated format that cannot be interpreted without the runtime
- Drop-in inference engine that replaces standard model loaders, requiring no code changes for deployment
- Embedded private key in the runtime ensures that original weight values are never exposed during execution
- Negligible performance overhead, delivering native inference speed on any hardware platform
- Protection against both GPU memory dumping attacks and direct file extraction, preventing weight theft and license evasion