NSA's "SKYNET" may be killing tens of thousands of innocent people


US National Security Agency(NSA) is using a machine learning algorithm "SKYNET" to identify "terrorists" by monitoring and analyzing Pakistan's mobile phone network, it is known from Edward Snowden leaked document I will. It is pointed out that such algorithms of SKYNET program have drawbacks, and thousands of innocent people may be classified as "extremists" in the US Government attack list.

The NSA's SKYNET program may be killing thousands of innocent people | Ars Technica UK

"Human Rights Data Analysis Group(HRDAG) 's data scientist Patrick Ball is a person who severely criticized the algorithm adopted by NSA's SKYNET as "bizarre and optimistic" or "completely false", and SKYNET is scientifically unfounded I talked to Ars Technica UK that it was enough.

Since 2004, 4,000 people have been killed by attacks by the US military drones, and most of the deceased were classified as "extremists" by the US government, the BritishBureau of Investigative JournalismIt reports. In 2014, former secretary of NSA and CIA said that "We are killing people based on metadata", NSA metadata analysis classification, the possibility that innocent people may be in danger It is whispering.

SKYNET collects metadata, stores information in NSA cloud server, extracts relevant information of terrorists, and identifies targets by machine learning. The collected data is gathered mainly from Dialed Number Recognition (DNR) data of call history of cellular phone, position information of the user, travel history, and the like, and is "turn off the mobile phone", "SIM When you take actions that try to avoid monitoring, such as "exchange cards", "exchange the terminal", it is also becoming a flag as a terrorist.

The collected metadata is analyzed with 80 properties that indicate the possibility of "terrorist" different from the general public, but what is most likely to be the terrorist in these machine learning programs is Osama · Successful exclusive interview with Al Qaeda leader such as Bin LadenAl JazeeraMr. Ahmad Muafak-Zidan, the chief of Islamabad's bureau chief.

SKYNET machine learning algorithmBayesian spam filterWe incorporate a re-learning program to accurately distinguish terrorists, like the data of "known terrorist" is supplied to the algorithm. However, it is a problem that the parameter of "known terrorist" is too small, and because it is unthinkable that an undiscovered terrorist will respond to the NSA's investigation, he said that it is insufficient as learning data for filtering .

In 2012, when SKYNET was built, the population of Pakistan was about 192 million and the population of mobile phones was about 120 million. The NSA analyzed about 55 million mobile phone population data, but it is thought that NSA relied on manual because there are too much data to analyze with only 80 variables. Therefore, it is thought that SKYNET could play only the role of assisting human judgment.

In addition, SKYNET sets a threshold classified as "terrorist" from the metadata and performs filtering. If the threshold is set as a false negative rate of 50%, half of the people will not be classified as "innocent" and will be classified as "terrorists". The threshold is set high because the real terrorist is not classified as innocent due to false detection. According to Mr. Ball, the threshold of 50% false negative rate is the same value as when making the final "homicide list", and the NSA lowers the threshold to prevent the missed rate of terrorists from increasing It seems not to be.

If you analyze the false negative rate at 50% against 55 million people, the false positive rates for which innocent people are classified as terrorists range from 0.008% to 0.18%. The number of people is about 15,000 people to about nine thousand people, so there is a possibility that innocent civilians are listed on the list as terrorists.

in Software, Posted by darkhorse_log