Modeling User Manipulation on Social Media

George T Amariucai

PITS Lab Research

  1. Privacy metrics for incomplete statistical information.
  2. Privacy in dynamic environments (time-varying features, social contexts, etc.).
  3. Security of cyber-physical systems (zero-dynamics stealthy attacks and privacy-utility tradeoffs to thwart them).
  4. Privacy strategies in online social networks.

User manipulation in social networks

How to define and quantify user manipulation

  1. First define a manipulation goal. Example: increase the probability that a user engages with (replies to, retweets, etc.) the manipulator’s messages.
  2. Measures a user's probability of engaging with a post in a neutral environment.
  3. Measures the degree of deviation of this probability  (manipulation gain) when the manipulator specifically targets this user:

– Manipulator optimizes its profile and message features to maximize the target user’s  probability of engagement.

User manipulation in social networks

How to defend against user manipulation

1.     Privacy-utility tradeoff:

  • User modifies the disclosed profile features (some directly, some through trained behaviors) to minimize (1) the probability of engaging with a message (think spam), or (2) the manipulation gain (think advertising).
  • User has to maintain a certain level of utility – e.g., the average trust level among his followers.

User manipulation in social networks

We can formulate a problem solvable by gradient algorithms.

1.     First build a function (regression) mapping manipulator features, message features and target features to probability of engagement.

User manipulation in social networks

3.    Finally, the outer optimization problem:

 

Note: We cannot formulate a proper minmax problem, as the true probability that the target user engages depends on target’s true attributes, not disclosed ones.

User manipulation in social networks

So how effective is such a manipulation strategy?

  • Apparently not very effective

Is there a way to manipulate better?

  • We can try a more complex manipulation campaign, consisting of multiple messages.
  • But we need a better user engagement model, mapping history to probability of engagement.

The Russian Troll Dataset

  • Released by Daren Linvill and Patrick

Warren (Clemson University) in July 2018

  • Published on GitHub by FiveThirtyEight, an American data-based news website
  • Contains 2,973,371 tweets from 2,848 unique Twitter handles (alleged trolls) connected to IRA (Internet Research Agency)
  • Tweets published from 2015 to 2017, collected with Social Studio

The Russian Troll Dataset

We want to study users’ interaction with the trolls; these interactions are still on Twitter.

  • Challenge: Troll tweets have been deleted from Twitter. Old tweets do not contain conversation ID, so not possible to retrieve complete conversation threads
  • Solution: We looked for tweets that constitute replies to troll handles; we fit them to troll tweets by matching the tweet times. We collected 48373 tweets from  41832 unique users from 2012-02-08 through 2018-05-30.

Spiking-neuron-based modelling

  1. Model each user as a spiking neuron
  2. Messages reaching the user act as inhibitory or excitatory post-synaptic
  3. Potential at the trigger zone P accumulates all the post-synaptic potentials (discounted as appropriate).
  4. The probability of firing (that is, of the user engaging with a certain message or topic) depends on the difference (P-T) between potential at trigger zone and some firing threshold 

Note: P and T are impossible to determine separately, and P amplitude is arbitrary, so we can set T=0.

Spiking-neuron-based modelling

Post-synaptic potentials can be defined in vector form:

  • One component (dimension) for each type of feature extracted from the exposure history.
  • Features have to be cumulative over the entire history.
  • Examples of history features:
    • Sentiment – must be decoupled into positive sentiment and negative sentiment
    • Popularity of the topics – for example, if a hashtag is present, popularity is the number of tweets, over the past week, that contained the hashtag.
    • Counts of words belonging to several word lists – for example, words eliciting hate, words eliciting pity, words eliciting outrage

Spiking-neuron-based modelling

Spiking-neuron-based modelling

Advantages of spiking-neuron-based modelling

  1. Realistic model of opinion formation
  2. Allows the study of:
    • Optimal complex manipulation strategies (e.g., phase 1 – trust build-up; phase 2 – manipulation attempt);
    • Manipulation strategies coordinated over a network of users, with multiple colluding trolls;
    • Various social media phenomena, like echo chambers, formation and mutation of ideas, etc.

A big “Thank You” to the graduate students who did this work:

Abiola Osho and Adaeze Okeukwu