metadata
license: mit
language:
- en
STEALTH
Secure Transformer for Encrypted Alignment of Latent Text Embeddings.
Model — short description
STEALTH is a 120M-parameter transformer encoder trained to produce encryption-invariant sentence embeddings. It learns a topology-preserving mapping from encrypted text embeddings to a plaintext embedding space using the Semantic Isomorphism Enforcement (SIE) multi-objective loss.
Model specs
- Architecture: 12-layer Transformer encoder with key-attentive attention and multi-key aggregation.
- Model size: ~120M parameters.
- Embedding dim (output): 256.
- Tokenizer: encryption-aware byte-level tokenizer.
Highlights
- ✅ Privacy-first: Operates on ciphertext without requiring decryption.
- ✅ Topology preserving: SIE loss aligns encrypted and plaintext embeddings while preserving semantic distances.
- ✅ Robust training: Multi-key augmentation (multiple ciphertext variants per plaintext) improves invariance and generalization.
- ✅ Practical: Small model footprint (120M) for efficient deployment in constrained environments.