Megadose AI progress, ranked and analyzed.

Google is making private AI practical with homomorphic encryption

· HN · Frontpage AI ·
Google says its HEIR compiler can turn trained AI models into versions that run inference on encrypted inputs.

HEIR is an open-source toolchain for homomorphic encryption, aimed at making private inference usable without a team of cryptographers. Google frames the benefit as letting servers compute on ciphertext and return encrypted results without seeing the underlying data. The post shows four demos: recommendations, credit-card fraud detection, encrypted network anomaly detection, and hotword detection. Latency figures are described as single-threaded CPU results, with accelerator work promised later. HN · Frontpage AI's note

score 5

Categories: Research

Discussions

  • hn · 261 points · 160 comments
  • hn · 271 points · 164 comments
  • hn · 278 points · 167 comments
  • hn · 287 points · 171 comments
  • hn · 296 points · 176 comments
  • hn · 300 points · 176 comments
  • hn · 308 points · 181 comments
  • hn · 316 points · 186 comments
  • hn · 324 points · 194 comments
  • hn · 337 points · 198 comments
  • hn · 344 points · 204 comments
  • hn · 354 points · 210 comments
  • hn · 361 points · 214 comments
  • hn · 368 points · 215 comments
  • hn · 378 points · 219 comments
  • hn · 381 points · 221 comments
  • hn · 388 points · 232 comments
  • hn · 394 points · 241 comments
  • hn · 404 points · 246 comments
  • hn · 410 points · 249 comments
  • hn · 418 points · 254 comments
  • hn · 425 points · 257 comments
  • hn · 432 points · 259 comments
  • hn · 439 points · 262 comments
  • hn · 443 points · 264 comments
  • hn · 448 points · 266 comments
  • hn · 453 points · 268 comments
  • hn · 458 points · 268 comments
  • hn · 458 points · 268 comments
  • hn · 463 points · 269 comments
  • hn · 465 points · 272 comments
  • hn · 468 points · 274 comments
  • hn · 472 points · 276 comments