So far, we have emphasized the shift from individual to group detectors, in an attempt to map the evolution of social bots. Now, let’s look at recent advances in the field to get a sense of the future of deception detection.
First, we observe that individual and group approaches to social bot detection follow a reactive pattern. In practice, when scientists and social media administrators identify a new group of accounts that behave incorrectly and cannot be effectively detected using existing methods, they react and start developing a new detection system. Therefore, the driving factor for developing new and better detectors has always been bot harm. The main consequence of this approach is that improvements in attacker detection usually occur only some time after evidence of new harm has been collected. Thus, attackers (bots, cyborgs, and trolls) benefit from the time it takes to design, develop, and deploy a new effective detector, during which they are essentially free to interfere in the online environment.
In other words, scientists are constantly one step behind the developers of malicious accounts.
This lag may explain the current state of our social ecosystems: despite the growing number of existing detection methods, the influence of bots and other bad actors on our online conversations seems to have not diminish.
The second observation concerns the use of machine mobile phone number data learning to detect social bots. The vast majority of machine learning algorithms are design to operate in stationary and neutral, if not safe, environments. When the assumptions of stationarity and neutrality are violat, the algorithms produce unreliable prictions, which lead to a sharp decline in performance [15].
Notably, the task of social bot detection is neither stationary nor neutral.
The stationarity assumption is violat by the mechanism of bot evolution, whereby accounts exhibit different behaviors and characteristics over time. Also, the neutrality assumption is the revenue cycle model – the catalyst you need for your marketing equations clearly violat, as bot developers actively try to fool detectors. As a result, the algorithms we have reli on, which have demonstrat excellent detection results in research, are actually severely limit in their chances of detecting bots in the wild. Developments in machine learning can come to the rescue and possibly mitigate both problems.
Adversarial machine learning is a paradigm specifically design for use in these scenarios, which represent adversaries interest in deceiving learn models [15]. Its high-level goal is to learn the vulnerabilities of existing systems and possible attacks to exploit them before these vulnerabilities are effectively exploit by attackers. Early detection of vulnerabilities, in turn, can facilitate the development of more robust detection systems.
One practical way to realize this vision is to create and experiment with adversarial network instances, i.e. input instances specifically design to induce errors in machine learning systems. All tasks relat to detecting online deception, manipulation, and automation are inherently adversarial. As such, they represent ripe areas for adversarial machine learning.
This intuition led to the first papers published in
2018–2019 that initiat the development of an adversarial approach to bot detection, as shown in Part B of Figure 5. In so-call adversarial bot detection, researchers experiment with meaningful adversarial examples with which they comprehensively test the capabilities of existing bot detectors [9]. In this context, adversarial examples can be complex types of existing bots and trolls that manage to evade detection european union phone number using current methods, or even bots that do not yet exist but whose behavior and characteristics are modeled, as done by Kreschi et al. [9], or bots designed specifically for experiments, as done by Grimm et al. [17].
Finding good examples of adversarial behavior can help scientists understand the weaknesses of existing bot detection systems. As a result, bot hunters will no longer have to wait for new malicious bot behaviors to adapt their methods. Instead, they can test them proactively (rather than reactively).