FOCI’21, August 27, 2021, Virtual Event, USA Padmanabhan et al. curfews to target organization of political dissent while minimizing the impact on the economy, as many sectors require Internet access during the day. Indeed, evidence suggests that the junta is aware of the economic impacts of censorship—Myanmar restored access to banking apps at the end of April coincident with the lifting of curfews, perhaps as a way to reinvigorate economic activity [76]. The evolution of censorship throughout this time period underscores the importance of being able to track multiple methods of censorship to gain a holistic understanding of the digital strategy of autocrats, something that has as of yet been difficult to do at scale [42]. While nightly outages have now ended, the social media and website blocking we described that has persisted since February 5 may indicate a move toward more selective methods of censorship [28]. This shift is consistent with a pattern in authoritarian regimes of engaging in targeted censorship to maximize political impact while minimizing its cost [5, 71]. We hope this paper can provide a template of combining Internet measurements to provide a broader understanding of digital strategy of autocrats, an effort that could be scaled and replicated cross-nationally in future work. Figure 5: Collateral damage to Twitter users outside Myanmar as a result of the BGP hijack event affecting Twitter’s address space. The figure shows Twitter traffic observed by Kentik from different source ASes (indicated by different colors) that was being routed towards Campana Mythic (AS136168). Our analysis of BGP data collected by the Routeviews and RIPE RIS projects shows the illegitimate route propagated (at least) to operators in Singapore (AS4844, AS56300, AS24482, AS132132) and Vietnam (AS45903) who received, accepted, and further propagated it. This resulted in collateral damage for Twitter users outside Myanmar. We quantify the extent of this collateral damage in Figure 5, which shows that a small volume of traffic from Kentik’s customers outside Myanmar was directed towards the hijacker AS136168 instead of AS13414 (Twitter). 4 5 CONCLUSION In this study, we used multiple complementary datasets to investigate the censorship events that occurred in Myanmar following the military coup on February 1 2021. These datasets revealed different facets of censorship: IODA data showed episodes of complete disconnection from the Internet with accurate timing, data from Kentik presented insights into cellular traffic restrictions, and OONI data demonstrated the blocking of social media and various websites. These datasets are complementary at various levels. One key difference is in their goals and design. OONI seeks to measure website/social media blocking, whereas IODA and the use of Kentik’s data target full connectivity disruption of Internet users. As such, they operate at different layers of the network stack and with different granularity. Though data from both IODA and Kentik can be used to measure full connectivity shutdowns, their measurements are distinct in nature and thus can each reveal unique insight on how disconnections affect different networks (e.g., IODA’s data can reveal outage timing patterns with more accuracy than Kentik’s, but IODA’s data sometimes lacks visibility into disconnections of cellular operators, whereas in this paper we show that Kentik’s data can be used to study such events). We believe that the lenses offered by these diverse datasets will be highly beneficial to analyses of future censorship events. Similar to Myanmar, recent Internet censorship efforts in other countries have also used a variety of censorship methods [62, 82, 83]. As censors evolve in their use of information controls, our ability to understand them will also need to develop. Thankfully, an increasing variety of open tools and datasets are being actively developed and deployed, enabling deeper and more timely visibility into network interference phenomena. DISCUSSION The censorship events in Myanmar reflect emerging patterns of politically inspired censorship and offer insight into the ways in which authoritarian regimes combine censorship approaches strategically to achieve their immediate goals. First, the timing of censorship during a coup is consistent with many studies that have shown that Internet censorship is targeted during sensitive political time periods and periods of potential power transitions, such as elections and large-scale protests [23, 25, 44, 47, 73]. Among many, recent examples of outages during political transitions have occurred in January 2021 in Uganda [82] and in the summer of 2020 during large-scale protests in Belarus [83]. The fact that the initial outages were implemented by the challenger rather than the incumbent government lends support to recent theoretical and empirical work that suggests that Internet censorship during a coup attempt can increase the probability of a successful coup [9]. Conspirators in a coup may benefit from shutting communications quickly, to prevent public or government coordination against their coup attempt [39]. Yet, as we mentioned before, the haphazard nature of the outages during the initial coup in Myanmar may reflect the difficulty of the challenger in implementing this censorship, and could be a reflection of their initial lack of political control. After consolidating power, the new junta in Myanmar began imposing Internet curfews, shutting down the Internet during the night while keeping it on in the day. While the imposition of nightly curfews has long been a tactic by authoritarian (and some democratic) regimes to quell protests, it has recently been adopted in the virtual world in the midst of large scale unrest. Recent examples of similar Internet curfews include Libya in 2011 [17] and Gabon in 2016 [13]. Like physical curfews, regimes may implement Internet Ethical considerations. We recognize that some of our results could be used by censors to implement more rigorous measures. However, since the majority of our analyses were derived from publicly available datasets, we believe that the benefits yielded by an empirical understanding of these events outweigh the risks [18, 49]. 32