STEALTH: Secure Transformer for Encrypted Alignment of Latent Text Embeddings via Semantic Isomorphism Enforcement (SIE) Loss Function

Abstract

The pervasive use of large language models (LLMs) on sensitive data presents a critical privacy challenge, as traditional encryption renders data unusable for inference. We introduce STEALTH, a 120M secure transformer framework designed to process encrypted text while preserving its semantic utility under an authorized-key threat model (no decryption or side-channel access). The core innovation of STEALTH is the Semantic Isomorphism Enforcement (SIE) loss function, a loss that trains the model to learn a topology-preserving mapping between encrypted text embeddings and their original plaintext latent space. This encourages preservation of semantic relationships and topological structure in the encrypted domain. Using retrieval-based reconstruction from a domain-aligned plaintext corpus, STEALTH achieves near-perfect semantic retrieval (BLEU score of 1.0 under full-corpus coverage in our experiments) and enables accurate privacy-preserving clustering on encrypted embeddings. We evaluate STEALTH across 44 datasets spanning general language understanding, healthcare, finance, legal, e-commerce, programming, content analysis, reading comprehension, and corporate communication domains with 16 encryption schemes (704 experimental conditions), establishing a comprehensive benchmark for privacy-preserving NLP on encrypted text. Performance depends on domain alignment between encrypted inputs and the indexed plaintext corpus. Our results demonstrate that, with well-aligned domain indexes and retrieval support, models can perform effective NLP on encrypted data without direct decryption.