embedding-inversion

Embedding Inversion via Conditional Masked Diffusion: recover original text from embedding vectors using parallel denoising. Live demo + training pipeline + technical report.

🛠 Reconstruct original text from text embeddings using conditional masked diffusion to reveal reversible embedding representations efficiently and accurately

✨ State-of-the-art privacy for multi-vector VLM retrievers (ColPali/ColQwen2) — a field-level PII linkage attack, the holographic leakage mechanism, and Cataract, an adaptively-verified index-time defense.

Domain-adapted OCR for multilingual medical documents: LoRA/PEFT adaptation of compact VLMs, seed-level reproducibility and an evaluation gate, the OCR error cascade under shortcut control, and redaction mechanisms (masking, adversarial gate, LEACE) under probe and inversion attacks. Paper code + Zenodo data (10.5281/zenodo.22078532).

Gradient-guided embedding inversion against BAAI/bge-m3 — proof that a stored embedding vector is not anonymized data.