I was the guy yesterday who posted on the forum about the spoofing myth of Niantic. What I basically said is that Niantic can implement a machine learning solution to detect the majority of spoofers and the refusal to do so implies a monetary incentive to keep cheaters in play. For those who don't know, a machine learning algorithm is not explicitly programmed, but observes the human behavior and adapts/evolves accordingly to do a task (e.g recommend a movie, a youtube video or in Niantic's case catch a fraud).
However, I was vehemently attacked by some wiseguys that this is not possible and started throwing me a bunch of technical nonsense like IP addresses/VPNs etc which of course are ineffective and that's why I didn't propose that. Some other guys proposed that AI is not the holy grail and innocent players may be banned too. That true, and can be overcomed by using an expert in AI. But the all-loving Niantic didn't hesitate to ban all rooted phones (and I think we can all agree rooting is not cheating).
So I tried to do a bit of research about Niantic and see myself what's the technical orientation of the employees. The best source to do that is Linkedin. I searched the profiles of all 108 employees at Niantic to see their technical skills https://www.linkedin.com/search/results/people/?facetCurrentCompany=10149011 . My suspicions were true. I didn't find a single person with skills in machine learning/AI/data science. This does not prove anything....yet.
Niantic a few weeks ago posted 3 positions for highly qualified PHDs machine learning engineers https://www.nianticlabs.com/jobs/machine-learning-engineer/. I am sure that at least one of them will be used in the field of fraud detection.
So now it makes sense. Niantic probably so far has not used any AI expertise to their advantage, and all the spoofers so far are banned because of network analysis, which is partly effective. I hope that these new employed data scientists start firing up Hadoop and Spark clusters in order to eradicate all frauds who ruin our game.
Yes they can do it. Is not that hard monitoring activity. Just use three variables. User, GPSCoordinates and Time. The GPSCoordinates can't vary in terms of kilometers in a short time. If the beheaviour gets repetitive even a human can analyze (post activity, taking gyms, gathering pokestops).
BUT this will reduce their outtome as thousands of kids are spoofers. Just check how many idio.ts got mad with Nicks reaction in YouTube Channel trainertips.. all of those fuckers are spoofers.
I don't rememebr anyone saying it can't be done...in fact I remember the discussion pointing out that a massive dent could be made in the 'sniping' in a far simpler and faster way...so yes we were in fact agreeing that it could be done if they chose to put a little more effort.
Putting that aside, it is good to hear that they are employing people whose job it may be to specifically do something about it. Better late than never.
Thanks for sharing what you found.
EDIT If you felt attacked, then please know that was far from my intention, and for fear of speaking for others, I didn't think anyone else was intentionally attacking you either. I just was more interested in the part where you were discussing the technical side of how it could be achieved.
Nice work and good initiative! I don't know much about coding but I feel like this is feasible with such a popular game. I think Niantic completely agrees that the game was not released in a completed state, and they are struggling to release more content to keep up a player base, while leaving spoofers on the backburner. We are complaining yes, but it is still such a popular game that is getting revenue and increases in player activity based mainly on events and releases of new content. Perhaps this is why it has been taking a while to fix.
I simply don't belive they will ever do it. They were supposedly looking for a people months ago and had advertised for Ai engineers but still no change. It is just a front to look like Niantic care.
I bet none of the spoofers even pay to play, why would you when you can teleport anywhere and take down gyms at your leasure.
I've only ever worked with genetic algorithms (GA) for solving engineering problems before so my knowledge as a whole on machine learning is quite limited. However, using GAs as an example, wouldn't you need to make your "fitness value" that that code would use to determine if the user is a cheater or not very high to ensure zero false positives? This would allow casual cheaters, those who only spoof occasionally, to continue unheeded.
Glad you give me the chance to share some technical background. Yeah, GAs are more common in tasks like creating agents and generally solving NP hard problems faster. What I propose is that a simple anomaly detection algorithm can cut out a great chunk of the spoofer userbase https://www.coursera.org/learn/machine-learning/lecture/Rkc5x/anomaly-detection-vs-supervised-learning. If you want to dive deeper, you can use a sophisticated neural network which can offer you tremendous accuracy. Neural networks have boundless capabilities. Google used them to beat the human champion in Go and in ten years we are promised self-driving cars.
Haha it's funny you should mention that, I am actually taking the Python Data Science course on Coursera now! I may actually be interested in taking that course after I finish what my current specialization. How would such an algorithm handle non-spoofers, who periodically behave abnormally? i.e. bouncing around due to weak GPS signal?
Good work! I strongly recommend to take Andrew Ng's course first! I also recommend this book for neural networks https://www.deeplearningbook.org/.
Spoofers (or bots) walk to a place to perform a particular action. An action is when you throw a ball, a berry, press on a gym etc and generally it means you send a client request to PoGo server. When you gps drift for example you may take an action eg spin a pokestop, but not always. Also the gps drift can't be more than 100 meters and your movements has a centroid. Those patterns are distinguishable. This is the most difficult case. If the fraud uses teleports then it's more easy to detect them
So you create a matrix for each user and store their info. You can create variables like the ratio of time or distance between 2 actions divided by number of pokemon caught or high IV pokemon caught. Spoofers usually use a tracker alongside, which will give anomalies in the variables aforementioned.
The truth is I cannot know 100% what variables Niantic stores. And the magic with machine learning is that you don't know the correlations initially. Apple for example found that users are more likely to buy Ipads than Iphones in summer. Sometimes probabilities cannot be explained and that's why you need a thorough research.
Each human player has some preferences. This in machine learning is called bias or overfitting. For example, I prefer to fight a gym which is located in the range of Starbucks, because I like the coffee there. An intelligent algorithm will capture that bias in a frequency vector.
Now assume someone makes an intelligent bot, which simulates human walking and visiting by injecting random numbers in an acceptable probability interval. Those actions won't indicate a bias, and will probably raise suspicions. Aside from that, no botter o r spoofer will enjoy cheating like this. They would like to make a snipe to catch that 100% Iv dratini or teleport to their nearest gym to reclaim the gym they just lost.
Now, what I said above is very useful. If Niantic captures my behaviour, they know that each time I visit a new city and I want coffee I will prefer Starbucks, so... that's a potential sponsor for Niantic! If they run analytics on their data they may find amazing correlations, who knows what, like for example 65% users in the noon who visit Walmart play Pokemo go!! So Niantic has a new sponsor!!
But if the majority of your userbase are spoofers, then you can't prove shit, because users don't go physically to the sponsor's place.
Very interesting, thank you for the for detailed responses! I was thinking along the same lines of using random numbers and probability intervals with rules. I'm sure there is a way to only travel on roads and randomly make turns or not. This definitely would not be ideal for a spoofer but could be beneficial for them just to experience/dust farm.
Techincally this might be not 100% accurate and an full algorithm is needed but just as a comment this can be done analyzing the distances and time.
I also inside a house get the crazy GPS but how far? It never goes more than 200 meters or 300 as way too much. We are talking of uses that travel 1km in the frame time of couple of minutes or even seconds. That's how it will work.
I kind of agree with what apax said. There are so many ways to detect spoofers. I once came up with a statistics method. You see, given # of total eggs hatched & # of total caught, the probability of having more than a certain amount of 100iv mons can be calculated (they must know the distribution). If it (so-called p-value) is < 0.01%, then they are 99.99% confident that the tested account has acquired mons in a non-random way.
So there must be some reasons that Niantic have been holding off banning spoofers. I'll discuss in another post (https://pokemongo.gamepress.gg/q-a/dilemma-niantic-banning-paying-spoofers-or-not), if you're interested please drop by and leave a comment.
I like this. One counter point though. If it became known that this was the methodology used to catch spoofers, then spoofers could simply start catching more low IV guys intentionally to fix their average IVs.
Edit:
Actually nevermind, I misspoke, you said just 100 IVs. That should work either way.
