Decoding Keypresses using Mobile Phone Accelerometers Woodams Clark Jesse Freeman Christopher Weeden Mobile Phones Mobile phones come with a wide array of sensors that allow them to interface with the world around them. This can also mean they can be used in unintended ways. Many OS now have users identify permissions for apps to use sensors - However many are still not tightly controlled Sensors in a SmartPhone Accelerometer Heart Rate Monitor Gyroscope Fingerprint Sensor Magnetometer Pedometer Proximity Sensor Light Sensor Barometer Accelerometer Cannot set permissions for apps to use accelerometer By using the accelerometer, vibrations can be recorded in order to decipher what a person is typing on their keyboard. Many studies have been done in the past using things that can record vibrations, or electromagnetic waves, but this equipment is obviously more sensitive and stronger than a standard consumer device Related Work ● ● ● ● ● Emanations of electrical and mechanical devices have been used to expose user’s activities as far back as 1940s TEMPEST - Referring to spying on information systems through leaking emanations, including unintentional radio or electrical signals, sounds and vibrations Techniques have been used to capture electromagnetic signals generated by teletype terminal in the 1940s. Used to detect Viet Cong trucks from distances up to 10 miles away. Techniques used on CRT and LCD screens in order to recover their contents. Using electromagnetic emanation Optical Emanation Optical emanation has been used to reveal contents displayed on CRT monitors by using diffuse reflections of the monitors. Work was later extended to capture reflections from a target’s eye. These attacks are effective at distances up to 30 meters, but are much more difficult to execute with an obstructed line of sight. Acoustic Emanations High technical devices are not needed when it comes to acoustic emanations. A consumer grade microphone will suffice. One can be used to figure out the contents of a dot-matrix printer with high accuracy. Requires the microphone to always be near the target. Attack and Threat Model In order for the previous types of attack to be possible, it must be easy to exploit the user’s surroundings. Using a phone increases the chance of attacks being possible. ● ● Come with a lot of sensors that can be exploited But can the sensors be used to actually get decipherable information and patterns from the target? Attack Model Continued Many users place their mobile devices on their desk when doing work. Model works as follows: ● ● ● adversary installs malicious application on the phone application will record accelerometer data by periodically sampling activity exfiltrate the collected data via connecting to the internet and uploading the data to a remote server Application of previous techniques There are two key problems associated with this work: ● ● Given that the accelerometers found in current mobile phones sample at rates that are orders of magnitude smaller than previous acoustic and electromagnetic attacks, can keypresses even be detected by these sensors? if such events can be observed can previously developed methodology such as identifying individual keys using neural networks be applied to identify keystrokes? Comparing two different mobile phones Comparing two different mobile phones cont. The iPhone 3GS provides very noisy output from the accelerometer. ● You can’t differentiate between specific key presses because of all the output The iPhone 4 has a much more exact accelerometer ● Gives a better reading for individual key presses What does this mean? Accelerometer-based eavesdropping could potentially be implemented in phones! With the accuracy of the iPhone4’s accelerometer there’s a potential to eavesdrop simply by putting the phone next to a keyboard Possibly with newer technology using an iPhone 6s there could be better results? The experiments set-up The researchers used the following materials to test the potential of eavesdropping: ● ● ● iPhone4 Wooden Desk Apple Blue-tooth keyboard Each key was then pressed 150 times to provide training data Single key press results Using machine learning algorithms to determine single key results the accuracy was very poor ● Achieved only a 25.89% accuracy with the iPhone4 ● This was only a third of the results of previous work 78.85% ● The iPhone’s accelerometer only samples data at a rate of 100Hz ● Previous work also had an accelerometer that sampled at a rate that was 441 times higher Acoustic accuracies vs Seismic Accuracies Due to the Implications Due to the extremely low sample rates of each key this required that a pair wise system be put in place to improve accuracy: ● Pi and Pj are two sequential keypresses The relation between the two keypresses have two features: ● ● Horizontal Orientation: The location of each keypress event relative to a central line that divides the keyboard in half Distance between Consecutive keypresses: for a threshold of distance in keys there is a distance where two keys are either classified as either “far” from each other or “near” to each other The Central line is defined as the line between the keys t,g,b (the left side) and y,h,n (the right side) The Word Canoe This two keypress scheme is used to identify more features of the word that was typed with relation to each of the keys pressed to narrow down the possibilities. Each word of length n will have n-1 pairs that are output the word canoe can be broken down into these pairs of keypresses: “ca” “an” “no” “oe” which would be represented as: LLN. LRF. RRF. RLF since “ca” are both on the left side of the keyboard and were classified as “near” each other the above results shows that it is represented as LLN Results L/R Distinguishing pairs that are L/R and N/F is easy, even by eye. Left/Right pairs have large GForce difference, whereas a Near/Far pair has minimal GForce difference N/F Results Left Ultimately this means that distinguishing individual keys is hard, but detecting “Regions” of keys is easy. L/R classifier was able to correctly identify 91% of the individual keypresses as right or left, and 70% of the keypress pairs as near or far. Right Results Conjunctions The frequency of correctly matched words declined greatly due to 2 and 3 letter words. In english, 2 and 3 letter words tend to be conjunctions (and, the, to, or, an…) and can be easily inferred when the rest of the sentence is decoded. Typed Text: The birch canoe slid on the smooth planks Recovered Text: *** punch canoe slid ** *** smooth planks Typed Text: Glue the sheet to the dark blue background Recovered Text: Glue *** sheet ** *** dark blue background Results Accuracy Experiment 1 First experiment attempted to decode one of the first 10 Harvard sentences, whereas the second experiment attempted to decode all 10. Both used a dictionary of all words in the first 10 Harvard sentences. The accuracy of first-word-choice was only 46%, much worse than 80% from experiment 1. However, when looking at second-word-choice accuracy rises to nearly 73% Experiment 2 Results - Comparison While the results don’t appear incredible, it’s important to keep perspective. ● ● Using accelerometers has a much lower sample frequency than most other methods such as Acoustic Dictionary Techniques proposed by Berger. Able to work on words of length 4+ characters ○ ● Bergers’ uses 7-13 and has best results with repetitive words Despite this, accuracy is still rather competitive to these more robust methods. Limitations The largest limitation to this method is distance, noise and surface material. The ability to distinguish keypresses from noise drops with respect to the Inverse Square Law. A phone more than 2 feet away cannot distinguish keystrokes from noise. People walking by, an air-conditioner running, or finger tapping can also completely disrupt the ability to decode keystrokes. And having the phone on any soft surface (napkin, cloth, phone-case) will also diminish acoustic capabilities. Conclusions While preventing this attack is trivial (keep phone away from keyboards), the premise attack exploits secondary sensors in mobile devices, which the user assumes trust of. These sensors are assumingly innocuous, but this paper shows they can be taken advantage of to attack the user. A better way to prevent these attacks is to ensure access to all sensors of a device are restricted in an appropriate manner, handled by the OS. References Research Paper Link: - http://www.cc.gatech.edu/fac/traynor/papers/traynor-ccs11.pdf