All projects

Secure Cancer Prognosis via fhe

Secure Cancer Prognosis
FHE Toolkit
11 mo

The challenge

The most important clinical process in patients with cancers is the accurate estimation of prognosis and survival duration. Patients who volunteer their genomic data run the risk of privacy invasion. While established encryption techniques, such as Advanced Encryption Standard (AES) can secure Personal Health Information (PHI) in Acquisition and Storage, they can only assure secure storage. Assuring data privacy in Computation is much more challenging.

The goal

View the live site

User Research

Read more
View the full case study

the process

Claim 1&2 :

The above solutions for survival prediction and analysis claim to be very effective approaches but cannot be applicable in the real world due to the privacy concerns associated with genetic data. Our solution uses a similar approach but performs computations on homomorphically encrypted data, eliminating out the privacy concerns associated with it.

2) This project [3] demonstrates the use of Homomorphic Encryption for outsourcing neural-network predictions in the case of Acute Lymphoid Leukemia (ALL). Their work proposes a Neural Network model which will be able to predict ALL with approximate 80% accuracy and by the use of FHE also tackles the privacy concerns.

Claim 3 :

Most of the prior arts have used FHE with conventional prediction algorithms like deep learning methods but in survival prediction, more prominent approaches are Survival analysis methods, which are explicitly designed to deal with data about terminal events where some of the observations can experience the event and others may not i.e censored observations hence we have incorporated Kaplan Meier curves for survival prediction with other algorithms for better results.

3) This study [4] takes a large set of genomic markers (SNPs) and determines which of them are associated with a given trait. They used the SIMD capabilities of the CKKS HE scheme, to compute a large number of models, demonstrating the use of the CKKS homomorphic encryption scheme to train logistic regression models.

Another paper [5] evaluates the feasibility of training neural networks on encrypted data. Their proposed system uses the open-source FHE toolkit HElib to implement a Stochastic Gradient Descent (SGD)-based training of a neural network. The paper adopted the approach of the lookup table by pre-computing loss function and activation function and performing homomorphic table lookup.

The solution in [6] can handle thousands of records and hundreds of fields which takes a few hours to run. To achieve this performance they had to be extremely frugal with expensive bootstrapping and data-movement operation, they also aim to serve as a guide for what is or is not currently feasible to do with fully-homomorphic encryption talking about how expensive FHE can be and use of  (Somewhat homomorphic encryption) SWHE scheme in some scenarios.

Claim 4 : 

These papers have demonstrated the feasibility of training models via homomorphically encrypted data with the challenges that come along with it. On reviewing them in our scenario training on encrypted data would lead to difficulty in feature extraction, expensive computation and high training time, whereas if we go through our approach of training on unencrypted data it would lead to valuable feature selection and a better-fine-tuned model.

Read more
View the full case study
No items found.


We propose a cryptographic solution to the above-mentioned problem: maintain patient privacy, by encrypting all genomic data in the database and  perform a meaningful computation on the homomorphically encrypted data to predict the survivability of the patients who underwent specific therapy.

Read the full case study
View the live site




Amrin Khan


Virendra Pawar


Saurabh Rane

Sakshi Oswal

Parth Pawar

Tarun Meditya

Mitanshu Bhoot