Research map / Large language models / Instruction and fine-tuning
AlienLM: Alienization of Language for API-Boundary Privacy in Black-Box LLMs
Jaehee Kim, Pilsung Kang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Alien Adaptation Training (AAT) on top of the vocabulary bijection (ablation) | AlienLM · 7-benchmark average | average accuracy: 24.77 -> 52.92 | Improved |
| use token-ID relabeling instead of character-level ROT13-ASCII substitution (ablation) | AlienLM · 7-benchmark average | average accuracy: 23.34 -> 52.92 | Improved |
| optimize the bijection with embedding similarity instead of random bijection (ablation) | AlienLM · 7-benchmark average | average accuracy: 36.96 -> 52.92 | Improved |
| optimize the bijection with embedding similarity instead of random bijection (ablation) | AlienLM · GSM8K | exact match: 31.08 -> 63.08 | Improved |
| use proxy-model embeddings rather than target-model embeddings for bijection optimization (ablation) | AlienLM · 7-benchmark average | average accuracy: 55.07 -> 53.32 | No clear effect |
| lower alienization ratio rho (permute fewer tokens) (ablation) | AlienLM · 7-benchmark average | average accuracy: Pearson r=-0.96; about 86% RR at rho=0.6 | Improved |
| exclude domain data from AAT (ablation) | AlienLM · GSM8K | accuracy: 55.5% -> 41.7% | Hurt |
| exclude domain data only from bijection optimization (-Tokenizer) (ablation) | AlienLM · General Avg / GSM8K | accuracy | No clear effect |
| add 150K extra domain examples to AAT (ablation) | AlienLM · GSM8K | accuracy: 55.5% -> 60.6% | Improved |
| train AAT exclusively on domain data (ablation) | AlienLM · general benchmarks | General Avg: 50.55 -> 44.15 | Hurt |
| use per-user bijection seeds instead of a shared seed under collaborative leakage (ablation) | AlienLM · Collaborative token recovery (1K pairs) | token recovery: 6.05% -> 4.40% | Improved |
| train a single tenant model on five mixed bijection seeds instead of per-key models (ablation) | AlienLM · 7-benchmark average | average accuracy: 50.95 -> 41.26 | Hurt |
Closest papers by meaning
- Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives (NeurIPS 2024 · similarity 0.89)
- PrivAuditor: Benchmarking Data Protection Vulnerabilities in LLM Adaptation Techniques (NeurIPS 2024 · similarity 0.87)
- SentinelLMs: Encrypted Input Adaptation and Fine-Tuning of Language Models for Private and Secure Inference (AAAI 2024 · similarity 0.87)
- Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory (ICLR 2024 · similarity 0.86)
- Split-and-Denoise: Protect large language model inference with local differential privacy (ICML 2024 · similarity 0.86)
- Efficient and Privacy-Preserving Soft Prompt Transfer for LLMs (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
