LoREnc: Low-Rank Encryption for Securing Foundation Models and LoRA Adapters

Researchers introduced LoREnc, a training-free method to protect foundation models and LoRA adapters from unauthorized recovery.

LoREnc combines spectral truncation, compensation, and orthogonal reparameterization to block model reconstruction while preserving legitimate outputs.

The framework adds less than 1% overhead and does not require retraining or access to original training data.

It offers a practical defense against model extraction and IP leakage for deployed generative AI systems.