I've Played a Part in the Formation of the Surveillance State

I was so naive back in 2007. It simply sounded like a cool thing to do for a thesis.

Funny how my perception of the world has changed throughout the years. I've become more of a consequentialist. I try to see the possible effects of my actions if I were to do A, B, or C. Thinking about it, did my work do harm more than good? The fruits of my thesis were never mine. It's the property of the university, even though I've toiled away 16 hours a day. There were hundreds of thousands of lines, in which I trimmed, trimmed and trimmed to make it superlite. I've worked so much that I even hear ghosts typing on the keyboard when I was sleeping.

News was that the university sold the piece of tech to FB and some other companies.

My work back then predates modern deep-learning algorithms that you so often hear today. But mine was especially lightweight - it revolves around an algorithm called SIFT - Scale Invariant Feature Transform. It recognizes faces, buildings, and just about anything with contrasting visual topology.

Back then I added two design goals using this SIFT algorithm.

First is to have a continuous readout of people in feeds - sort of like building a 3d model of that person. SIFT was able to detect gait, clothing features, and such. Whenever the person enters the frame again, SIFT was able to detect them even without their faces - thanks to the signature representation of the built 3d models stored in the database.

Second is to optimize SIFT for low-resolution videos. My experiments were conducted with only low-res crappy phones, webcams and CCTVs. My university cafe area has about an average of 800-1200 people per day, in and out. After running the test for about a week, my results were at 81% accuracy within acceptable confidence. People with visually featureless faces (too fair or too dark) were picked up as well thanks to the the 3d models built.

I wonder if any governments have been implementing that. But one of the news that I've gotten is that NSA, and BIG-DATA in general just have too much data that it often turns into noise. Curve-fitting and categorizations has become extremely difficult. Deep learning for surveillance purposes is going to be a mountainous challenge. And that's because data is becoming more HD, and police states are trying to crosslink so much data and try to guess the possible type of crimes. It's not so much about tracking people only when they need to.

And that's why I think certain simplicities will never go away. Maybe modern surveillance states should stay stubborn in trying to implement "smarter" algorithms that churn out probabilistic numbers on potential criminal activities. It's more like a META-type pattern recognition now, instead of just pure visual pattern recognition.

I think that META-strategy would just crash and burn.