AI is increasingly built and used across trust domains: data owners train models together, private data meets remote frontier models, and agents delegate to other agents.
We study how each party keeps control, and we ship the tools.
| Project | Layer | What it does |
|---|---|---|
Figaro |
🟣 data | Federated learning experiments from natural language: an agent plans, runs, and compares reproducible runs. |
FedArena |
🟣 data | Attack/defense arena for federated learning: submit a method by prompt or code, benchmark it against baselines, rank it. |
CloakBot |
🟠 capability | Local privacy kernel for frontier LLMs. A local Gemma 4 model sanitizes text, images, and documents before any request leaves your machine; values are restored locally. |
PyFE4AI |
🟢 foundations | Functional encryption for trustworthy AI: 42 inner-product and quadratic schemes, adapters for federated aggregation and encrypted inference. |
Research directions
| Direction | In one line | Topics |
|---|---|---|
| Data & models | Data stays put; capability flows. | federated and continual learning for foundation models · controllable secure aggregation · poisoning and backdoor defense · machine unlearning |
| Agents | Agents may act on their own; their permissions stay bounded. | privacy for prompts sent to frontier LLMs · agent memory and delegation security · multi-agent risk evaluation |
| Foundations | AI to engineer cryptography; cryptography to protect AI. | functional and homomorphic encryption · TEE-assisted confidential aggregation · cryptographic tooling for ML |
How we work
Local first · measured, not asserted · reproducible · open. Sensitive data stays on its owner's side; every privacy claim ships with an evaluation; configs, seeds, and artifacts make results replayable; prototypes stay open to inspect and extend.
SPIRE takes its letters from Security, Privacy, and Trustworthiness · questions or collaboration: open an issue in any of our repositories