Computing on data no one has to reveal.
OBLIV Lab — short for oblivious, the cryptographic property of computing without seeing — builds cryptographic protocols and systems that let organisations analyse, train on, and transact with sensitive data while it stays secret — even when a majority of participants are dishonest.
Four directions, one principle: privacy by construction.
Secure multi-party computation
Efficient protocols for comparison, aggregation, and evaluation in the dishonest-majority model — the setting where most parties may collude.
Private AI and machine learning
Training and inference on data that never leaves its owner: private RLHF, privacy-preserving prediction, and secure activation functions.
Robust federated learning
Aggregation that survives malicious majorities, poisoning, and inference attacks — without a trusted server.
Secure and private payments
Anonymous task assignment, worker payment, and verifiable transactions for crowdsensing and digital finance.
Research that runs.
Each line of work ends in a working system with measured throughput and communication cost.
LTS+
Secure comparison for distributed data under a malicious majority. Two- to threefold throughput gain and up to 3× lower communication than prior work, with direct use in ReLU and federated aggregation.
Obliv-FW
Distributed, oblivious firewall built on Bloom filters and secret sharing. Rules are split across servers, so no single machine learns the policy, and updates need no downtime.
SVEvote
An e-voting design combining secret sharing, distributed hashing, and MAC verification to deliver ballot privacy, integrity, and public verifiability without central trust.
Looking for partners in the US, Canada, and Australia.
We co-author, co-supervise, and co-build. If you work on MPC, private ML, or federated learning and want a committed collaborator, let's talk.