SecTL: Secure and Verifiable Transfer Learning-based inference
Résumé
This paper investigates the possibility of realizing complex machine learning tasks over encrypted inputs with guaranteed integrity. Our approach combines Fully Homomorphic Encryption (FHE) and Verifiable Computing (VC) to achieve these properties. To workaround the practical difficulties when using these techniques - high computational cost for FHE and limited expressivity for VC, we leverage on transfer learning as a mean to (legitimately) decrease the footprint of encrypted domain calculations without jeopardizing the target security properties. In that sense, our approach demonstrates that scaling confidential and verifiable encrypted domain calculations to complex machine learning functions does not necessarily require scaling these techniques to the evaluation of large models. We furthermore demonstrate the practicality of our approach on an image classification task.
Origine | Fichiers produits par l'(les) auteur(s) |
---|---|
Licence |