100 100 80 ToD-K Top-K (a)Non-victims'interferences (b)Interference of different motions (c)Interference at different distances Figure 17.Keystroke recognition top-k accuracy with different levels of interference from non-victims. different interfere distances.A volunteer plays the role of a 50.9% victim to perform keystrokes at a distance of 10 meters away from the receive antenna,other volunteers are treated as non- 9.19 victims in the target area.An illustration of the experiment is shown in Figure 17(a). In the first experiment,non-victims were requested to Test Data perform different movements within 5 meters of the victim. (a)Different victims Top-K ac-(b)Different training sets accur- including sitting,standing,and walking.Figure 17(b)shows curacy acy the top-K accuracy of the keystroke recognition under the Figure 18.Keystroke recognition with different victims and the training set. above interferences.We observe that as the intensity of non- victims'actions increase,the recognition accuracy decreases significantly.It is worth noting that:first,there is no significant back of the victim.The volunteers always typed with their impact on the accuracy of recognition when someone is sitting right hand and the keyboard was always placed in the right still,even if there are multiple non-victims around.Second, front of the volunteer during the experiments.From Figure the standing posture has more significant influence on the 16(b)we can observe that the performance of SpiderMon is performance than the sitting posture,because humans move consistent for the front,back,and right orientations,while the the body involuntarily even when standing still.Third,the performance on the left is considerably worse.This could be impact of walking on the signal is so significant that the caused by the occlusion of the typing hand(right hand)by the keystroke action is completely submerged. victim's moving body when viewed from the left In the second experiment,non-victims were asked to main- Impact of Different Victims:We evaluate the impact of tain the sitting or walking state within different distances from different typing styles with three volunteers as the victims. the victim.The impact of these interferences are shown in The evaluation is based on the single keystroke setup.In Figure Figure 17(c).We observe that a sitting person has nearly no 18(a),we show the keystroke recognition accuracy of the three effect on keystroke recognition,even if it is within 2 meters of participants when training by his/her own data,where V1,V2, the victim.The walking action,even at a distance of 20 meters, and V3 represent three different victims.We observe that while still has an intensity higher than the keystroke action,the top-1 the top-1 accuracies for the three victims are different,i.e., accuracy rate is only about 25%,and the top-3 accuracy rate 94%,62%,and 78%,all victims'top-3 accuracies are over is less than 60%,barely better than a random guess. 95%.We further evaluated the performance when the training and testing data are from different victims (one victim's data VII.CONCLUSIONS as the training set and another victim's as the testing set).The In this paper,we show that LTE reference signals can be top-1 results of the accuracy are shown in Figure 18(b).In used as a medium for side-channel attacks by implementing Figure 18(b),the digits in each grid mean the top-1 accuracy the SpiderMon system that displays and analyzes LTE CRS when the testing data is from Va and the training data is from signals in real-time.Compared to previous attacks that use Wi- V.and the diagonal data represents the accuracy of using Fi CSI,LTE-based attacks can achieve comparable perform- his/her own data with 10-fold cross-validation.We observe that ance while have a longer operational distance and do not need when using different people's data for training,the accuracy is active transmissions.Therefore,LTE-based attacks are harder significantly reduced.For example,for V1,when the training to be detected and lead to more serious security breaches.We data is from V2 and V3,the accuracy drops from 94.0% hope that our work could inspire more research in this area to to 52.0%and 48.8%.However,we believe this problem can protect users from such attacks. be alleviated by collecting more people's keystroke data and training with a more powerful machine learning algorithm that ACKNOWLEDGMENT is less sensitive to the variance of users,e.g.,with a GAN [45]. We would like to thank our anonymous shepherd and re- Non-victims'interference:To evaluate the performance viewers for their valuable comments.This work is partially when other non-victims are in the target area,we conducted supported by National Natural Science Foundation of China two sets of single keystroke recognition experiments concern- under Numbers 61872173,61872174,61972192,and Collab- ing the interference of different movement intensities and orative Innovation Center of Novel Software Technology.Rx 10m (a) Non-victims’ interferences 1 2 3 4 5 6 7 8 910 Top-K 0 20 40 60 80 100 Recognition Accuracy (% ) w/o interference sitting sitting (2 person) standing walking (b) Interference of different motions 1 2 3 4 5 6 7 8 910 Top-K 0 20 40 60 80 100 Recognition Accuracy (% ) w/o interference sit 2m sit 5m sit 10m walk 5m walk 10m walk 20m (c) Interference at different distances Figure 17. Keystroke recognition top-k accuracy with different levels of interference from non-victims. V1 V2 V3 0 20 40 60 80 100 Recognition Accuracy (%) top 1 top 2 top 3 (a) Different victims Top-K ac￾curacy 94.0% 51.5% 50.9% 52.0% 62.0% 49.1% 48.8% 50.4% 78.0% V1 V2 V3 Test Data V1 V2 V3 Training Data 40 50 60 70 80 90 100 (b) Different training sets accur￾acy Figure 18. Keystroke recognition with different victims and the training set. back of the victim. The volunteers always typed with their right hand and the keyboard was always placed in the right front of the volunteer during the experiments. From Figure 16(b) we can observe that the performance of SpiderMon is consistent for the front, back, and right orientations, while the performance on the left is considerably worse. This could be caused by the occlusion of the typing hand (right hand) by the victim’s moving body when viewed from the left. Impact of Different Victims: We evaluate the impact of different typing styles with three volunteers as the victims. The evaluation is based on the single keystroke setup. In Figure 18(a), we show the keystroke recognition accuracy of the three participants when training by his/her own data, where V1, V2, and V3 represent three different victims. We observe that while the top-1 accuracies for the three victims are different, i.e., 94%, 62%, and 78%, all victims’ top-3 accuracies are over 95%. We further evaluated the performance when the training and testing data are from different victims (one victim’s data as the training set and another victim’s as the testing set). The top-1 results of the accuracy are shown in Figure 18(b). In Figure 18(b), the digits in each grid mean the top-1 accuracy when the testing data is from Va and the training data is from Vb, and the diagonal data represents the accuracy of using his/her own data with 10-fold cross-validation. We observe that when using different people’s data for training, the accuracy is significantly reduced. For example, for V1, when the training data is from V2 and V3, the accuracy drops from 94.0% to 52.0% and 48.8%. However, we believe this problem can be alleviated by collecting more people’s keystroke data and training with a more powerful machine learning algorithm that is less sensitive to the variance of users, e.g., with a GAN [45]. Non-victims’ interference: To evaluate the performance when other non-victims are in the target area, we conducted two sets of single keystroke recognition experiments concern￾ing the interference of different movement intensities and different interfere distances. A volunteer plays the role of a victim to perform keystrokes at a distance of 10 meters away from the receive antenna, other volunteers are treated as non￾victims in the target area. An illustration of the experiment is shown in Figure 17(a). In the first experiment, non-victims were requested to perform different movements within 5 meters of the victim, including sitting, standing, and walking. Figure 17(b) shows the top-K accuracy of the keystroke recognition under the above interferences. We observe that as the intensity of non￾victims’ actions increase, the recognition accuracy decreases significantly. It is worth noting that: first, there is no significant impact on the accuracy of recognition when someone is sitting still, even if there are multiple non-victims around. Second, the standing posture has more significant influence on the performance than the sitting posture, because humans move the body involuntarily even when standing still. Third, the impact of walking on the signal is so significant that the keystroke action is completely submerged. In the second experiment, non-victims were asked to main￾tain the sitting or walking state within different distances from the victim. The impact of these interferences are shown in Figure 17(c). We observe that a sitting person has nearly no effect on keystroke recognition, even if it is within 2 meters of the victim. The walking action, even at a distance of 20 meters, still has an intensity higher than the keystroke action, the top-1 accuracy rate is only about 25%, and the top-3 accuracy rate is less than 60%, barely better than a random guess. VII. CONCLUSIONS In this paper, we show that LTE reference signals can be used as a medium for side-channel attacks by implementing the SpiderMon system that displays and analyzes LTE CRS signals in real-time. Compared to previous attacks that use Wi￾Fi CSI, LTE-based attacks can achieve comparable perform￾ance while have a longer operational distance and do not need active transmissions. Therefore, LTE-based attacks are harder to be detected and lead to more serious security breaches. We hope that our work could inspire more research in this area to protect users from such attacks. ACKNOWLEDGMENT We would like to thank our anonymous shepherd and re￾viewers for their valuable comments. This work is partially supported by National Natural Science Foundation of China under Numbers 61872173, 61872174, 61972192, and Collab￾orative Innovation Center of Novel Software Technology