EavesDroid: Keystroke Recovery using Smartphone Accelerometers

In this paper, we demonstrate the usage of a smartphone accelerometer to eavesdrop on a nearby computer user and recover text based on recorded keyboard vibrations. We present EavesDroid, a proof-of-concept application that can be used to record and process accelerometer data and leverage it to reconstruct the original typed text. EavesDroid uses AdaBoost with Decision Stumps and achieves up to 85% accuracy in recovering original words with an expected error rate of 2 letters per word. We also make our dataset of accelerometer recordings available to the public, the first such dataset to the best of our knowledge.

##Workflow

• Generate Features: Follow the steps in code/scripts/README.md to generate the default labeled features file data/features/all-letters-labeled.csv for the whole of raw data
• Generate Training/Test Set: Divide the features into training (66%) and testing set (33%) using
  $ cd code/feature_extractor
$ java GenTrainingTesting
  Generates the training/testing files at
• data/dataset/training.lr.arff,
• data/dataset/training.ud.arff
• data/dataset/testing.csv
• Prediction Model Building: Run the following .R script to build the prediction model
  $ cd code/boosting
$ Rscript adaboost-lr-ud.R
  It creates the models at the locations
• code/boosting/adaboost.rf.lr.rda
• code/boosting/adaboost.rf.ud.rda
• Predicting: In order to get predictions on sample text, say data/dictionary/harvard1.txt, the following steps need to be followed

1. Automated Feature Generation: Generating features for the simulated signals of the corresponding words from the file data/dataset/testing.csv
$ cd data/tester
$ java GenParagraphArff < data/dictionary/harvard1.txt
   This generates the un-labeled features for all the words in the given text at the location data/paragraph_arff
2. Making Predictions: Use the following commands to generate the word predictions for the unseen words
   $ cd code/scripts
$ features-to-prediction.sh paragraph_arff 72 5 true
   This makes word predictions using 72 dictionaries and gives all the predicted words (at data/possible-words) with up to a Hamming distance of 5 from the label predicted for the given simulated word signal.
3. Accuracy: Accuracy of the predictions can be done using
   $ cd code/tester
$ java Accuracy