Sameer Wagh — privacy researcher and founder of SecretBit Ventures. Applied cryptography, secure computation, and privacy-preserving machine learning.
You have 2ⁿ visually identical coins, each weighing 0 or 1 gram. In at most n weighings on a balance scale, you can always determine whether the count of zero-weight coins is even or odd. The key is a binary search over a swap path.
The Stained Glass Transform is a transformer-based network trained to scramble token embeddings while preserving their utility under LLM inference. The output is somewhat magical: the model still produces coherent output but the original prompt is unrecoverable.
Sixteen cards, four binary properties, and an LLM agent left alone with a machine and its weekends -- chasing D(16), the fewest probes that always pin down a hidden arrangement. It walked the best-known strategy down 64 → 60 → 49 → 47 → 46, one probe above the information floor, where an old ghost from sorting is waiting.
Fast private lookup tables, explained through one idea you already know — Parseval's theorem. A dot product is the same in the signal domain and the wavelet domain, so compute it wherever it's cheapest.
Evaluating a frontier model on sensitive data when neither side wants to trust the other — work with Anthropic and the UK AI Safety Institute, on the OpenMined blog.
Using card-based cryptography to show how two people can compute the AND of their private answers using nothing but a handful of playing cards, and why that little trick is the seed of an entire field called secure multiparty computation.
Piranha was the first to train a neural network end-to-end under secure computation. The hard part wasn't speed — it was knowing why a run failed when you can't see a single value.
The Schwartz–Zippel lemma is the workhorse behind countless protocols — over fields. Pika carries it to rings, where zero divisors get in the way, and leaves a crisp open question: how small can the slack constant be?
For years, secure multi-party computation chased fewer rounds of communication. FALCON asked what happens at ML scale — and the answer flips the optimization target from network to silicon.
Private ML normally juggles two secret-sharing types — arithmetic for linear layers, Boolean for the rest — bridged by a costly conversion. SecureNN does the hard parts with just + and ×, and skips the bridge entirely.