New Encryption Method Shields Sensitive Data Sent to Cloud AI Services

When users interact with large language models through cloud APIs, their sensitive prompts, responses, and training data travel unencrypted over external systems—a significant privacy vulnerability. A new research paper introduces AlienLM, a privacy-focused technique that scrambles text before sending it to third-party AI services while maintaining the model's ability to process and respond to queries.

How AlienLM Works#

The approach relies on vocabulary-scale bijection, a mathematical transformation that maps each word in a language to a unique substitute within an "Alien Language." This encoded version can be sent safely to external API providers, and clients retain the original mapping needed to decode responses without the service provider ever seeing the plaintext.

Rather than requiring changes to the model itself, AlienLM uses standard fine-tuning to teach existing AI systems to operate directly on alienized inputs—a technique the researchers call Alien Adaptation Training (AAT). This allows deployment via API-only access, without needing model source code or special infrastructure.

Performance and Security Results#

Testing across four different LLM backbones and seven performance benchmarks, AlienLM preserved over 81% of the original model's accuracy on average. The method substantially outperformed simpler approaches like random token substitution and character-level encoding.

Even under adversarial attack—where attackers had access to model weights, statistical information from training corpora, and learning-based decryption attempts—the encoding proved resilient. Adversaries could only recover fewer than 0.22% of alienized tokens, demonstrating meaningful privacy gains against sophisticated threats.

Practical Implications#

The research suggests a viable path toward privacy-conscious deployment of cloud-based AI systems. By reducing plaintext exposure at the API boundary while sustaining task performance, the method addresses a growing concern as organizations increasingly rely on external LLM providers with sensitive data.