Abstract
Near-Field Communication (NFC) is a cornerstone of modern contactless payments but remains fundamentally vulnerable to relay attacks. While previous research suggested that ambient sensors could verify device proximity by comparing environmental "fingerprints," these studies often used unrealistic recording durations. This reproducibility study evaluates 17 sensors under a strict 500ms transaction window. Using a large-scale dataset of over 1,000 transactions across four environments, the study demonstrates that even advanced deep learning models cannot achieve the security thresholds necessary for commercial payment systems.
The Threat: NFC Relay Attacks
NFC technology is susceptible to relay attacks, where an adversary "bridges" the physical gap between a victim's device and a legitimate payment terminal. In a typical scenario, one attacker device emulates a terminal to the victim, while another emulates the victim's phone to the actual terminal, potentially miles away. The victim authorizes what they believe is a local transaction, while their credentials are being relayed to authorize a fraudulent purchase elsewhere.
Proposed Solution: Ambient Sensing
To counter these attacks, researchers proposed using ambient sensors (e.g., light, sound, atmospheric pressure) to create a unique environmental signature. If both the phone and terminal record matching signatures simultaneously, they are presumed to be co-located. This study investigates whether this theory holds up when moving from controlled lab settings to noisy, fast-paced real-world environments.
Methodology and Realistic Constraints
The authors developed a custom Android test-bed to evaluate 17 different sensors. Unlike prior work, this study enforced a 500ms recording window, aligning with EMV industry standards for contactless payment performance. The evaluation framework included:
- Similarity Metrics: Statistical comparisons like Mean Absolute Error (MAE) and Correlation Coefficients.
- Traditional Machine Learning: Random Forests and Support Vector Machines (SVMs).
- Deep Learning: Artificial Neural Networks (ANNs), Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs).
The study involved 252 participants across four distinct locations (lab, cafeteria, dining hall, and library), ensuring the dataset captured the "noise" of real-world interactions—tapping, holding, or hovering the phone—within a realistic timeframe.
Key Findings: The "500ms Wall"
The most significant discovery was the impact of the 500ms constraint. Many sensors, including GPS, Bluetooth, and WiFi, exhibited failure rates exceeding 90% when restricted to such a short sampling window.
Atmospheric pressure emerged as the most reliable individual sensor, likely because it is less susceptible to interference from the user's hand compared to light or sound sensors. However, even the best-performing models—such as a CNN combining magnetic, light, and motion data—achieved an Equal Error Rate (EER) of only 0.246. For comparison, a Random Forest model using pressure data achieved an EER of approximately 0.10.
Conclusion
While ambient sensing is a compelling theoretical defense, this study highlights a critical gap between lab results and real-world requirements. An EER of 0.10 to 0.25 is far below the security standards required for financial transactions, where biometric standards often target a False Acceptance Rate of 1 in 50,000. Future research may need to explore pre-emptive recording or wait for higher hardware sampling rates to overcome the "500ms wall."
Summary updated