Reproduce “AlienLM: Alienization of Language for API-Boundary Privacy in Black-Box LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jaehee Kim, Pilsung Kang
AlienLM introduces a deployable method for enhancing API-boundary privacy in black-box LLMs by translating plaintext into an uninterpretable "Alien Language" through a secret, vocabulary-level bijective permutation. It adapts target LLMs via API-only fine-tuning, consistently retaining over 81% of original task performance across multiple benchmarks and models while demonstrating empirical robustness against various recovery attacks.
Key insights
- LLMs can effectively learn and operate in a cryptographically transformed "Alien Language" solely through API-only fine-tuning, even with a vocabulary-scale bijection.
- Optimizing the bijection using proxy embeddings from open-source models is sufficient to maintain semantic relationships critical for task performance under black-box constraints.
- A controllable 'alienization ratio' allows fine-grained balancing between the degree of plaintext exposure reduction and the preservation of task utility.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.