ml-driven smartphone privacy attack via power analytics

The rapid growth of smartphone usage has sparked a proliferation of public phone-charging hubs to cater to peoples’ growing charging needs. By virtue of being located in public spaces, however, these hubs have the potential to be manipulated by malicious actors who seek to use them as a vehicle to launch cyberattacks against the users of these hubs. In this paper, we show that if such a public charging hub is rigged with a power measurement circuitry, the power measurements could enable the inference of videos watched by the user on the phone. Using a playlist of 100 YouTube music videos, we show this kind of attack to classify the videos with an accuracy of up to 94.49%. We rigorously examine the dynamics of the attack using 5 different phone models, 16 screen brightness and volume configurations, and various training configurations and show it to be highly effective under a wide range of settings. Depending on the content of such videos, the profile of the target, and the attacker’s aims (e.g., government vs. private hacker), we argue that such an attack could have far-reaching privacy implications. The paper adds to the body of work highlighting power side-channels on computing devices as a potent threat to user privacy.

This paper designs and evaluates this line of attack. Specifically, we make the following contributions:

1) Studying the color underpinnings of the attack:

At the lowest level, a music video can be cast as a combi- nation of three color channels (R, G, and B) whose relative intensities in each pixel location vary over time. As a starting point towards understanding the origins of this line of attack, we undertook a proof-of-concept experiment to investigate whether each of the R, G, and B colors imprint a unique pattern on the power measurements made within the charging hub. To this end, we created three videos, each of which had a single color, and studied their power patterns. Our observations revealed the three videos to occupy three fairly distinct clusters, providing us with the first evidence that inherent differences in the brightness dynamics of each of the R, G, and B colors translate into separable patterns in the power drawn at the charger. Evidence from this ex- periment strongly suggested that more sophisticated videos (with time-varying colors and moving objects) might even be more uniquely identifiable and paved the way for our fully-fledged experiments on a large corpus of videos.

2) Evaluation of attack based on 100 music videos:

Following the promise depicted by our RGB experiment, we extended our investigations to 100 fully-fledged music videos. The videos were drawn from the BillBoard Hot 100 website from two timelines: week of June 29th, 2019 and June 27th, 2020. Using a catalog of fifty time and frequency domain features extracted from the charging power, we classified these videos with up to 94.49% test accuracy.

3) Sensitivity analysis of attack behavior:

To further understand the dynamics of the attack, we played music videos at sixteen different audio volumes and screen bright- ness levels to observe their effects on the inference of music videos. Variations in these attributes enabled us to evaluate the threat posed to users having different viewing habits. Results observed using Random Forest and Extra Trees classifiers revealed that the attack success is driven by the brightness of the screen and weakens with the decrease in the phone brightness. However, the impact on the attack with the changing volume levels does not reveal a distinct pattern. We proceeded to execute this attack on a variety of Android phone models: Google Pixel 2XL, OnePlus 5, OnePlus 5T, Samsung Galaxy S8+, and Sam- sung Galaxy S9 under extensive training configurations. We found that the attack performs best when phone models match between train and test datasets but still performs reasonably when phone models differ across training and testing datasets.