Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Safe Crime Prediction: Encrypted Deep Learning for Less Intrusive Surveillance

iamtrask.github.io

Post reply on HN
Print view View original

31–40 of 54 posts

Re: Safe Crime Prediction: Encrypted Deep Learning for Less Intrusive Surveillance

#31

Post by rrggrr »

williamtrask wrote: post #14
sambull wrote: post #13

"A positive prediction should launch an investigation, not put someone behind bars directly." That is the scariest thing I've read. All we need is a black box to investigate people at anytime.

There are already thousands of black boxes. Consider satellite photography, malware detectors, fire alarms, sniffing dogs, credit fraud detectors. All of these are tools ("black boxes") used to launch investigations. The future of crime fighting is impossible without tools like these as criminals become more sophisticated. The real question is who owns them and are they audit-able? This blogpost is about making neura…

Agreed. Stop, assess and arrest is now audit-able thanks to body cameras and we're already seeing the results in greater accountability. However, with greater visual accountability has come de-policing, a serious problem in some cities.

Re: Safe Crime Prediction: Encrypted Deep Learning for Less Intrusive Surveillance

#32

Post by ja30278 »

TTPrograms wrote: post #18

The number of people killed by terrorism in developed countries per year is, what, a hundredth of those killed in car accidents? A tenth of those shot in Chicago alone? http://www.businessinsider.com/death-risk-statistics-terrori... Terrorists have to kill literally hundreds of times more people per year for it to justify even the meagerest response. And one may argue about the economic impacts, but that's just a ref…

Counterpoint: As Taleb points out, car accident deaths are 'thin-tailed'; that is, the rate of car fatalities is essentially fixed (and predictable by looking at past data), while terrorism is 'fat-tailed', and the number of deaths that would be caused by a dirty-bomb in Manhattan does not appear in historical data, and so it can't be 'priced' in the same way.

Re: Safe Crime Prediction: Encrypted Deep Learning for Less Intrusive Surveillance

#33

Post by TTPrograms »

williamtrask wrote: post #26
TTPrograms wrote: post #18

The number of people killed by terrorism in developed countries per year is, what, a hundredth of those killed in car accidents? A tenth of those shot in Chicago alone? http://www.businessinsider.com/death-risk-statistics-terrori... Terrorists have to kill literally hundreds of times more people per year for it to justify even the meagerest response. And one may argue about the economic impacts, but that's just a ref…

Author here. So the techniques in this post aren't actually limited to terrorism, although it' s probably the easiest to talk about. It's broad enough to include homicides (as mentioned in the post), and hundreds of thousands of people are murdered each year because they aren't able to predict their own danger soon enough for law enforcement to be informed (and to intervene).

I should say - that's not to say that the technical content isn't very interesting. I just think that the solutions to violence are better considered from the perspective of preventing people from becoming violent people, rather than preventing violence from occurring. I think there's significant evidence to suggest that there are very powerful yet unused paths forward in that direction.

There are other applications that such a system could be interesting for, though. What about determining interest rates on loans for people and small businesses? Small businesses could benefit greatly if lower-rate small loans enabled them to grow. Maybe insurance rates too? Etc.

Re: Safe Crime Prediction: Encrypted Deep Learning for Less Intrusive Surveillance

#34

Post by TTPrograms »

ja30278 wrote: post #32
TTPrograms wrote: post #18

The number of people killed by terrorism in developed countries per year is, what, a hundredth of those killed in car accidents? A tenth of those shot in Chicago alone? http://www.businessinsider.com/death-risk-statistics-terrori... Terrorists have to kill literally hundreds of times more people per year for it to justify even the meagerest response. And one may argue about the economic impacts, but that's just a ref…

Counterpoint: As Taleb points out, car accident deaths are 'thin-tailed'; that is, the rate of car fatalities is essentially fixed (and predictable by looking at past data), while terrorism is 'fat-tailed', and the number of deaths that would be caused by a dirty-bomb in Manhattan does not appear in historical data, and so it can't be 'priced' in the same way.

In the example of the dirty bomb, I'm not sure that the number of casualties/cancer incidences is especially high, based on some brief search. Sept 11 seems like it was a more damaging attack, for example, and that seems roughly as difficult to replicate as a dirty bomb. Especially given the trade-off between payload and shielding your substances enough to prevent detection by the myriad of radiation detection systems I'm sure DHS has installed in major cities.

But you are certainly correct that statistically these events are more difficult to interpret, and raw averages are not a fair comparison.

Re: Safe Crime Prediction: Encrypted Deep Learning for Less Intrusive Surveillance

#36

Post by grandalf »

Does anyone know what minimum level of "homomorphism" is necessary to insulate the entity doing surveillance or analytics from legal action?

I ask this about the following contexts:

- government surveillance: does using a homomorphism remove the need for a search warrant?

- in-app analytics: does using homomorphisms allow a firm to consider data not disclaimed by the firm's privacy policy?

- research: when does a homomorphism eliminate the need for IRB approval?

Re: Safe Crime Prediction: Encrypted Deep Learning for Less Intrusive Surveillance

#37

Post by williamtrask »

grandalf wrote: post #36

Does anyone know what minimum level of "homomorphism" is necessary to insulate the entity doing surveillance or analytics from legal action? I ask this about the following contexts: - government surveillance: does using a homomorphism remove the need for a search warrant? - in-app analytics: does using homomorphisms allow a firm to consider data not disclaimed by the firm's privacy policy? - research: when does a hom…

Great question. I sincerely doubt there is legal precedent in this area though. I'd love to be wrong here though :)

Re: Safe Crime Prediction: Encrypted Deep Learning for Less Intrusive Surveillance

#38

Post by AceOfSpace »

I'm concerned with a tool such as this that can be directed to target anything of interest to those administering the system. We've already seen mass surveillance shift from being used "exclusively" for terrorism to also assisting in the war on drugs. Where else would the lens be trained? As the author suggests, murder. What about organized crime? Gangs? Illegal downloads ... the sharing of Netflix passwords?

Re: Safe Crime Prediction: Encrypted Deep Learning for Less Intrusive Surveillance

#39

Post by darawk »

TTPrograms wrote: post #18

The number of people killed by terrorism in developed countries per year is, what, a hundredth of those killed in car accidents? A tenth of those shot in Chicago alone? http://www.businessinsider.com/death-risk-statistics-terrori... Terrorists have to kill literally hundreds of times more people per year for it to justify even the meagerest response. And one may argue about the economic impacts, but that's just a ref…

> and terrorists would stop doing it, since it would be less effective at achieving the subgoal of "terror".

And that's why there are no terrorist attacks in Israel.

Re: Safe Crime Prediction: Encrypted Deep Learning for Less Intrusive Surveillance

#40

Post by CiPHPerCoder »

There are multiple elements to unpack here. The obvious pre-crime problem other commenters raise. The fact that terrorism is such a rare risk it isn't worthy of our attention.

But I'd like to raise another objection: Homomorphic encryption does not provide integrity over the ciphertext, which could open the door to active attacks against the systems that undermine its privacy goals.

https://news.ycombinator.com/item?id=14443191

https://paragonie.com/blog/2016/08/crypto-misnomers-zero-kno...

If you really need to build such a dangerous and needless system, would you want it to be built with such an error-prone cryptographic design? I'd say "No".

Post reply on HN