Privacy as a property you can prove.

Secure multi-party computation

Most MPC protocols assume an honest majority. Real deployments rarely offer that luxury. We focus on the dishonest-majority model, building primitives — secure comparison, oblivious evaluation, distributed policy enforcement — that remain both secure and practical when most parties may collude.

Representative work: LTS+ (secure comparison with EdaBits and a novel reduction), Obliv-FW (oblivious distributed firewalls), secure outsourced virtual network functions, privacy-preserving watermarking.

Secret sharingOblivious evaluationDishonest majority

Private AI and machine learning

Modern AI is trained on data its owners would rather not share: health records, conversations, location traces. We build training and inference pipelines where that data stays encrypted or secret-shared end to end, including privacy-preserving reinforcement learning from AI feedback, private traffic prediction, and secure evaluation of non-linear layers such as ReLU.

We are especially interested in making private fine-tuning and private inference of large language models affordable.

Private inferenceRLAIFLLM privacy

Robust federated learning

Federated learning removes the need to centralise data, but not the need to trust the participants. Our Fortified Federated Learning line designs aggregation that tolerates a malicious majority of clients and a curious server, defending against model poisoning and gradient-inference attacks at the same time.

Secure aggregationByzantine robustnessPoisoning defence

Secure and private payments

Payments leak who paid whom, for what, and when. We study anonymous task assignment and worker payment for mobile crowdsensing, verifiable e-voting as a transaction system, and privacy-preserving recommendation for sharing economies — all settings where money or value moves between parties who should not learn each other's identities.

Anonymous paymentVerifiabilityCrowdsensing