/enri-thumbnails/careeropportunities1f0caf1c-a12d-479c-be7c-3c04e085c617.tmb-mega-menu.jpg?Culture=en&sfvrsn=d7261e3b_1)
/cradle-thumbnails/research-capabilities1516d0ba63aa44f0b4ee77a8c05263b2.tmb-mega-menu.jpg?Culture=en&sfvrsn=1bc94f8_1)
7e6fdc03-9018-4d08-9a98-8a21acbc37ba.tmb-mega-menu.jpg?Culture=en&sfvrsn=7deaf618_1)
HintSight Technology
Synopsis
HintSight Technology introduces the invention Hybrid PP-NN, a privacy-preserving AI solution enabling encrypted data processing with applications such as in medical reports. It achieves state-of-the-art accuracy in less than a second for tasks like facial recognition.
Opportunity
AI-as-a-Service (AIaaS) has become a crucial trend supporting the growth of the digital economy. Digital service providers use vast amount of customer data to train AI models for image recognition, financial modelling, pandemic modelling and more, offering these as services on the cloud. While third-party models provide numerous benefits, uploading data to the cloud raises significant privacy concerns, especially with increasingly stringent privacy regulations and legislation.
To promote the adoption of AIaaS while addressing privacy issues, HintSight Technology proposes an advanced technology based on fully homomorphic encryption (FHE). This technology enables the sharing of data and AI models in a privacy-preserving manner, removing privacy hurdles when adopting third-party AI models. This approach allows existing neural networks to process encrypted data and produce encrypted outputs, which are only accessible by the model users, within an operationally acceptable time (e.g., less than one second for facial recognition in border control systems). Experimental results show that practical tasks such as facial recognition and text classification achieve state-of-the-art inference accuracy in under one second.
Technology
HintSight Technology provides an effective FHE-based solution for privacy-preserving neural network (PP-NN) inferences over encrypted data. The PP-NN protects clients’ privacy in AIaaS. In FHE-based PP-NN, the client encrypts sensitive data before sending it to the server. The server evaluates the neural network over encrypted data, produces an encrypted output, and returns the ciphertext to the client. Only the client, who has the private key, can decrypt the result. The server does not have the private key and cannot decrypt the input or output.
Accuracy and inference latency are critical for the practical use of PP-NN solutions. HintSight Technology focuses on both encryption algorithms and model structure:
- A customised fully homomorphic encryption is proposed, along with an efficient design for non-linear function evaluation.
- A new model, hybrid PP-NN, splits a deep neural network into a plaintext evaluation part and a ciphertext evaluation part.
- This allows for facial recognition to be performed in one second with only a 1MB encrypted message from clients to the server, whereas previous FHE-based PP-NN models would take more than one day.
Figure 1: Privacy-preserving neural network and its application in medical reports
Figure 2: We propose a new model - hybrid privacy-preserving neural network (Hybrid PP-NN) which splits a deep neural network into a plaintext evaluation part and a ciphertext evaluation part.
Applications & Advantages
Main application areas include facial recognition/verification, speaker verification, object classification, text classification, logistic regression and decision tree.
Advantages:
Enables the sharing of data and AI models in a privacy-preserving manner, removing privacy hurdles when adopting third-party AI models.
Allows existing neural networks to process encrypted data; produce encrypted outputs, which are only accessible by the model users.
Allows for facial recognition to be performed in one second with only a 1 MB encrypted message from clients to the server, whereas previous FHE-based PP-NN models would take more than one day.
