GraphTrack: A Graph-Based Mostly Cross-Device Tracking Framework
Cross-gadget tracking has drawn rising attention from each commercial companies and the general public because of its privateness implications and functions for person profiling, personalized services, and many others. One particular, huge-used kind of cross-machine monitoring is to leverage looking histories of consumer devices, e.g., characterized by a listing of IP addresses utilized by the units and domains visited by the gadgets. However, current shopping history based mostly strategies have three drawbacks. First, they cannot seize latent correlations among IPs and domains. Second, their performance degrades considerably when labeled system pairs are unavailable. Lastly, they don't seem to be strong to uncertainties in linking looking histories to units. We suggest GraphTrack, a graph-based mostly cross-machine tracking framework, to track users across completely different units by correlating their shopping histories. Specifically, we suggest to mannequin the advanced interplays amongst IPs, domains, and devices as graphs and seize the latent correlations between IPs and between domains. We assemble graphs that are sturdy to uncertainties in linking searching histories to gadgets.
Moreover, we adapt random stroll with restart to compute similarity scores between gadgets primarily based on the graphs. GraphTrack leverages the similarity scores to carry out cross-device monitoring. GraphTrack doesn't require labeled gadget pairs and luggage tracking device might incorporate them if accessible. We evaluate GraphTrack on two real-world datasets, i.e., a publicly out there mobile-desktop monitoring dataset (round 100 users) and a a number of-device monitoring dataset (154K users) we collected. Our results show that GraphTrack considerably outperforms the state-of-the-art on both datasets. ACM Reference Format: Binghui Wang, Tianchen Zhou, Song Li, Yinzhi Cao, Neil Gong. 2022. GraphTrack: A Graph-based mostly Cross-Device Tracking Framework. In Proceedings of the 2022 ACM Asia Conference on Computer and Communications Security (ASIA CCS ’22), May 30-June 3, 2022, Nagasaki, Japan. ACM, New York, NY, USA, 15 pages. Cross-gadget monitoring-a method used to determine whether numerous gadgets, resembling mobile phones and desktops, have frequent homeowners-has drawn a lot consideration of both commercial corporations and the general public. For example, Drawbridge (dra, 2017), an promoting firm, goes beyond traditional gadget tracking to determine units belonging to the same user.
Because of the rising demand for cross-system monitoring and corresponding privacy considerations, the U.S. Federal Trade Commission hosted a workshop (Commission, 2015) in 2015 and released a employees report (Commission, 2017) about cross-machine monitoring and business laws in early 2017. The rising interest in cross-gadget tracking is highlighted by the privateness implications associated with monitoring and the functions of luggage tracking device for user profiling, personalized providers, iTagPro and iTagPro device person authentication. For example, a bank application can adopt cross-device tracking as part of multi-factor authentication to extend account security. Generally talking, iTagPro bluetooth tracker cross-device monitoring mainly leverages cross-device IDs, background setting, or looking historical past of the units. For example, cross-gadget IDs could embody a user’s e-mail deal with or username, which are not relevant when customers don't register accounts or do not login. Background atmosphere (e.g., ultrasound (Mavroudis et al., 2017)) additionally cannot be utilized when units are used in numerous environments resembling home and workplace.
Specifically, looking historical past primarily based monitoring utilizes source and destination pairs-e.g., the client IP address and the vacation spot website’s area-of users’ searching records to correlate totally different devices of the same person. Several searching history primarily based cross-gadget tracking strategies (Cao et al., 2015; Zimmeck et al., 2017; Malloy et al., 2017) have been proposed. For instance, IPFootprint (Cao et al., 2015) makes use of supervised learning to investigate the IPs generally used by gadgets. Zimmeck et al. (Zimmeck et al., 2017) proposed a supervised methodology that achieves state-of-the-art efficiency. Particularly, their method computes a similarity score through Bhattacharyya coefficient (Wang and Pu, 2013) for a pair of devices primarily based on the frequent IPs and/or domains visited by both devices. Then, they use the similarity scores to trace devices. We call the method BAT-SU because it makes use of the Bhattacharyya coefficient, where the suffix "-SU" indicates that the method is supervised. DeviceGraph (Malloy et al., 2017) is an unsupervised methodology that models devices as a graph primarily based on their IP colocations (an edge is created between two gadgets if they used the same IP) and applies group detection for monitoring, i.e., the devices in a neighborhood of the graph belong to a person.