A Labeled Inner-Product Functional Encryption Framework with Client-Side Vector Transformation for Secure Inference

Abstract

 Inner-Product Functional Encryption (IPFE) enables functional evaluations directly over encrypted data, making it a viable primitive for privacy-preserving machine learning inference. However, standard deployments face two operational vulnerabilities: structural key-misuse resulting from unauthorized key-ciphertext pairings, and direct input reconstruction attacks from legitimately decrypted functional outputs. To address these vulnerabilities simultaneously, this paper presents a unified dual-layer defense framework. The primary layer incorporates a label-based cryptographic binding mechanism (Labeled-FHIPE) to enforce context consistency between functional keys and ciphertexts under the Decisional Diffie-Hellman (DDH) assumption. The secondary layer applies a client-side linear transformation matrix  prior to encryption, introducing a higher-dimensional structural obfuscation to protect original vector elements from post-decryption inference. We validate the framework's theoretical properties through an empirical implementation using Python. The experimental evaluation demonstrates that mismatched key-label queries are deterministically restricted (0% execution rate), while adversary reconstruction error for original inputs scales with the expansion ratio . The results indicate that integrating cryptographic access controls with spatial input obfuscation provides a practical layer of security for semi-honest inference environments.