AlienLM: Alienization of Language for Privacy-Preserving API Interaction with LLMs

ICLR 2026 Conference Submission25556 Authors

Keywords: Encryption, Obsfucation, LLMs

Abstract: We introduce $\textbf{\textit{AlienLM}}$, a framework that reinterprets encryption as language translation for large language models accessed exclusively through black-box APIs. Existing approaches based on secure inference or differential privacy and federated learning offer limited protection in API-only scenarios. $\textbf{\textit{AlienLM}}$constructs an Alien Language through a vocabulary-level bijection and employs API-only fine-tuning, thereby ensuring compatibility with commercial black-box services while requiring no access to model internals. Across four LLMs and seven benchmarks, $\textbf{\textit{AlienLM}}$ preserves more than 81\% of the original performance, substantially surpasses substitution- and obfuscation-based baselines, and exhibits strong robustness against token-mapping and frequency-analysis attacks. $\textbf{\textit{AlienLM}}$ provides a deployable, low-overhead mechanism for safeguarding sensitive data in API-mediated applications such as healthcare, finance, and education. More broadly, our findings reveal a practical separation between linguistic representation and task competence, thereby motivating future work on composable privacy-preserving layers and formal characterizations of the learnability–opacity trade-off.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 25556