Show HN: Locket – Robust feature-level access control for LLMs (github.com)

Locket, an innovative feature-locking technique (FLoTE) presented by researchers Lipeng He, Vasisht Duddu, and N. Asokan, aims to enhance the flexibility of large language models (LLMs) by enabling pay-to-unlock schemes for specific features. This approach is significant for the AI/ML community as it allows developers to monetize their models more effectively while offering users tailored access to functions based on their needs. This could lead to innovative business models and broaden the applications of LLMs across various industries. Technical details reveal that Locket utilizes Low-Rank Adaptation (LoRA) to lock features in models such as DeepSeek-Math-7B, with experiments conducted using high-performance hardware (8 × NVIDIA A100 GPUs). Key components include the ability to train specific LoRA adapters for single and multi-feature scalability, which can adapt components of models based on user interaction or payment. The paper also outlines configuration settings for training and evaluating robustness against various attack methods, reflecting its potential applications in creating both secure and monetizable AI models.