Introducing the Urban Anonymity Simulation

I had a weird experience on the way into work this morning. Standing on the platform for the Q-Line, I watched the flow of traffic with the realization that every driver on Woodward could see my face, and yet I felt completely unseen. And not in a bad way.

Cities oftentimes make me feel this way. An exposed public space should feel like a stage (and I, a performer), but if that’s the case then I share the stage with an endless procession of co-stars. I sometimes feel less like the center of the stage than a passive observer of someone else’s play. City residents compromise and change roles at will: one is often the performer, and other times we simply observe. There is an interplay of anonymity and observability in such spaces.

This isn’t new to urban scientists, who describe the sort of blasé attitude of city-dwellers who have seen too much and begin to restrict their attentions. But this notion of anonymity extends beyond the city to arbitrarily large groups of individuals. Latanya Sweeney abstracts the notion of the crows to a dataset where each row is an individual observation. She then asks, how many additional datapoints are necessary to hide the influence of the one? At what point does the crowd become big enough such that anonymity is assured?

The answer is complex, and we get into it in our literature review of k-anonymity. But the more interesting question is how this notion of anonymity evolves when our presence in a city leaves data traces at every interaction (card swipe, bus ride, phone unlock via facial recognition). In such a case, the “crowd” in which a person anonymizes themselves is surgically dissected until the urban anonymity of the past is infeasible.

We build a simulation that explores the question analytically. With the Urban Anonymity simulation, users discover their discoverability within a crowd of individuals. When we account for linked systems and salient distinguishable attributes, the physical crowd reduces to a much more refined cloud of our most immeadiately-observed neighbors.

Explore the standalone simulation: Urban Anonymity Simulation.

We argue that digital systems can erode or redistribute that anonymity without the subject’s consent. We use Sweeney’s notion of K-Anonymity to analyze the city as a dataset where the influence/exposure of an individual diminishes as the city grows larger. Digital systems nondeterministically replicate and proliferate this data based on the strength of the individual’s influence to train models and autonomous agents. But this means that an individual with less visibility is more replicable than one with more.

Privacy is not a matter of personal preference, or even personal control. It evolves from macro-conditions that emerge from a system of many interoperating subjects and infrastructures. Some spaces erode privacy more than others, and digital connectivity often contributes.

The question of urban anonymity isn’t a new one, but takes on new meaning when the physical space (and crowd) is digitized. Such a crowd is much smaller, with fewer places to hide. And while cities become more responsive to their residents, the definition of urban anonymity needs a new definition.

• Sensors transitioned from transporting data to performing those decisions on the device, and so a single decision-making body is split between a million of the city’s sensors.
• Asymmetric visibility produces a new form of urban subalternization since heavily-analyzed communities may not even be the beneficiaries of the model (ie. using surveillance data from low-income neighborhoods to train surveillance models serving high-income areas; using bus ridership data to train a model improving the routing for premium-class Uber trips)

We take this anonymity seriously, and see any feasible urban future as one built from data sovereignty of the producers. This fuels our experiments into data deletion, unlearning in optimizers, and the diffusion of information in networks. We redefine the notion of crowds and space with the digital in mind, as well as the physical. This gives us the tools (analytical and technical) to describe how our algorithms learn from our cities.