TL;DR
This paper introduces a novel side-channel attack that uses the acoustic emanations from PIN entry keypresses to recover PINs with 60% accuracy. As a countermeasure, it proposes a behavioral biometric system that leverages these same acoustic signals to verify user identity with 88% accuracy, enhancing the security of PIN entry devices.
What Problem Does This Paper Address?
In today's interconnected world, especially with the rise of the Internet of Things (IoT), secure user authentication is paramount. Personal Identification Numbers (PINs) are a common and convenient method for verifying user identity on resource-constrained devices like ATMs and Point-of-Sale (POS) terminals. However, these systems are vulnerable to various attacks. This paper focuses on a specific vulnerability: the acoustic emanations produced when users type their PINs. These subtle sounds, often overlooked, can be captured and analyzed to infer the entered PIN, posing a significant security risk. The paper aims to demonstrate this vulnerability and propose a practical, low-cost method to mitigate it by using the same acoustic signals for user verification.
Methods
The research presents a two-pronged approach: an attack and a defense.
Attack (PIN Recovery): The core of the attack lies in analyzing the acoustic signals generated by keystrokes on a PIN entry device (PED). The methodology involves:
- Data Collection: Recording audio of users entering PINs using a standard microphone.
- Feature Extraction: Identifying keypress events and extracting features such as press/release times, volume, and spectral characteristics. A crucial aspect is analyzing the time intervals (latency) between keystrokes.
- Time-Frequency Analysis: Employing techniques like cross-correlation and frequency-based distance to quantify the similarity between acoustic signals.
- Machine Learning: Training models (specifically Logistic Regression in the attack phase) on these extracted features to predict the entered PIN. The attack focuses on recovering individual digits of the PIN.
Defense (User Verification): The proposed countermeasure leverages the unique typing patterns of individuals, treating them as a behavioral biometric.
- Feature Extraction: Similar to the attack, inter-keystroke timings (latency) and keypress hold times are extracted.
- Machine Learning: Multiple machine learning models (Gaussian Naïve Bayes, Logistic Regression, and Support Vector Machines) are trained on these features to classify users. The goal is to distinguish between authorized users and imposters based on their distinct typing rhythms.
Datasets
The paper describes experiments conducted using a custom dataset.
- Data Collection: Acoustic data was collected from 15 users, each entering four distinct 4-6 digit PINs multiple times. The PINs were chosen to include various transitions (short, long, diagonal).
- Benchmark: While not explicitly stated as a new benchmark, the collected dataset serves as the basis for evaluating both the attack and verification models. The authors emphasize that their approach is designed to be generalizable to any modern PED.
Key Insights and Ideas
- Acoustic Emanations as a Rich Source of Information: The paper highlights that the seemingly simple sounds of keypresses contain enough information to both infer a PIN and uniquely identify a user.
- Inter-Keystroke Latency as a Behavioral Fingerprint: The timing between keystrokes, rather than just the sound of individual presses, is a strong indicator of user identity. Different users exhibit distinct patterns in how quickly or slowly they transition between keys.
- Low-Cost Attack and Defense: The methods rely on readily available hardware (microphones) and standard machine learning techniques, making the attack feasible and the defense cost-effective to implement.
- Unified Approach: The same acoustic signals that pose a security threat can be repurposed as a robust defense mechanism, creating a synergistic security solution.
Key Results
Attack: * The proposed PIN recovery attack achieved a 60% chance to recover the PIN key. * The attack model can predict single characters of a PIN sequence with an accuracy above 62%.
Verification: * The user verification model achieved an 88% accuracy in identifying the user. * Logistic Regression with latency and hold features yielded the best performance, achieving an 88% True Acceptance Rate (TAR) with a 12% False Acceptance Rate (FAR) and 12% False Rejection Rate (FRR).
Limitations and Future Work
- Training Set Size: The authors acknowledge that the training set size (15 users, limited PIN entries per user) is a limitation. They plan to increase the number of PIN entries per user (e.g., to 8 or 10 times) in future work to improve robustness.
- Generalizability: While the approach is designed to be general, the current experiments are based on a specific PED. Further testing on a wider variety of PEDs would be beneficial.
- Adversarial Sophistication: The current attack assumes a relatively straightforward recording and analysis. More sophisticated adversaries might employ advanced signal processing or noise reduction techniques.
- Future Work: The authors suggest extending this technique to other areas like remote authentication (website logins) and mobile phone unlocking, especially in scenarios where visual biometrics like Face ID are not feasible (e.g., due to mask-wearing). They also plan to explore different supervised machine learning models.
Summary updated