Latest / MBA Training Data / Privacy-enhancing technologies in practice: the hype is ahead of the implementation
Transcript
- 0:00Welcome to the Leaders Insights podcast.
- 0:07Today's episode, Privacy Enhancing Technologies in Practice.
- 0:11The hype is ahead of the implementation.
- 0:14Everyone's boardroom is buzzing about privacy enhancing technologies.
- 0:18Tools that let you compute on data without exposing it.
- 0:22You've deployed them.
- 0:23Tell me the uncomfortable part.
- 0:25The uncomfortable part is that most of what gets called a deployment is a demo that
- 0:29survived a friendly audience.
- 0:32The science is genuinely sound.
- 0:34The production reality is that the gap between a slide and a running system is measured in
- 0:39years and headcount you don't have.
- 0:42That sounds like the usual enterprise is slow excuse.
- 0:45Be specific.
- 0:46Take fully homomorphic encryption, computing on encrypted data without ever decrypting it.
- 0:52Beautiful math.
- 0:54In practice, it runs thousands to millions of times slower than plain computation.
- 0:59IBM and Zama have made real strides.
- 1:02But when a vendor tells you it's production ready, ask them to run your fraud model on
- 1:06it and report the latency.
- 1:08I've watched a query that took 40 milliseconds balloon to 15 minutes.
- 1:12So the technology just doesn't work?
- 1:15It works.
- 1:16It's just priced wrong for most jobs.
- 1:18That's the thing nobody says out loud.
- 1:20You deploy encrypted computation to keep a medical genome private.
- 1:24Worth every second.
- 1:26You would not deploy it to check whether someone clicked an ad.
- 1:29What about the safer sounding ones?
- 1:31Differential privacy, adding statistical noise so no single person can be identified.
- 1:36The US Census used it.
- 1:38They did in 2020.
- 1:40And it's the best cautionary tale we have.
- 1:43The census added mathematical noise to protect individuals.
- 1:46And demographers screamed because small town population counts came out visibly wrong.
- 1:51That's the trade nobody frames honestly.
- 1:54Privacy and accuracy are a dial, not a switch.
- 1:57Open up protection.
- 1:59Your numbers get blurrier.
- 2:00Every vendor pitch skips the part where your analysts lose precision they were relying
- 2:04on.
- 2:05You're a chief data officer.
- 2:07Aren't you supposed to be selling this internally?
- 2:09I'm supposed to not get fired for a breach or a lawsuit, which is a different job.
- 2:14And I'll tell you the dirty secret.
- 2:16Half the privacy problems teams are trying to solve with exotic cryptography can be solved
- 2:20by deleting data they never needed to collect.
- 2:23Radical, I know.
- 2:25That's the funny because it's true line.
- 2:27But give me a case where the fancy stuff actually paid off.
- 2:31Synthetic data, artificial records generated to mimic real ones statistically without
- 2:36being real people.
- 2:38American Express and several banks use it to test systems without touching live customer
- 2:43data.
- 2:44That's a genuine win because the accuracy bar for does my software crash is low.
- 2:51Where it fails is when someone trains a credit model on synthetic data and acts surprised
- 2:56when it misses the rare cases that actually cause losses.
- 3:00So the rule is match the tool to how much error you can eat.
- 3:04Exactly.
- 3:05And here's the number that should scare people.
- 3:07Gartner, a research firm so treated as directional, has said for a while that a large share of
- 3:12privacy tech pilots never reach production.
- 3:16My own experience puts real production deployment in the single digits of the companies claiming
- 3:20they have it.
- 3:22The rest have a proof of concept and a press release.
- 3:25Name what everyone gets wrong in one sentence.
- 3:29Everyone treats these as products you buy and switch on.
- 3:32They're research programs you staff with cryptographers and performance engineers.
- 3:36And if you don't have those people, you have a very expensive science fair.
- 3:40What about secure multi-party computation?
- 3:43Several organizations jointly analyzing data without any of them seeing the others raw
- 3:47records?
- 3:49Scientists love that pitch for fraud rings.
- 3:51They do.
- 3:52And the technology genuinely works for narrow agreed upon questions.
- 3:56The killer isn't the math.
- 3:58It's the lawyers.
- 4:00Getting three competitors to sign off on exactly which computation runs takes 18 months
- 4:04before a single line executes.
- 4:07The bottleneck is legal and governance, not silicon.
- 4:11Nobody budgets for that.
- 4:12So someone's listening, a pilot on their desk, bored breathing down their neck.
- 4:16What do they do, Monday?
- 4:18Take your single highest risk data flow, the one that would end up on the news.
- 4:23And ask one question.
- 4:24What's the largest accuracy loss or slowest speed I can tolerate for that specific job?
- 4:30Write the number down first.
- 4:32Then and only then let a vendor near you.
- 4:34If they won't test against your number, they're selling you a demo.
- 4:38Done.
- 4:39Want to know where you actually stand?
- 4:40Take the CDO self-assessment at MBA-training.com.