We Can Hear Your PIN Drop: An Acoustic Side-Channel Attack on ATM PIN Pads

Abstract

Personal Identification Numbers (PINs) are the most common user authentication method for in-person banking transactions at ATMs. The US Federal Reserve reported that, in 2018, PINs secured 31.4 billion transactions in the US, with an overall worth of US$ 1.19 trillion.

One well-known attack type involves the use of cameras to spy on the ATM PIN pad during PIN entry. Countermeasures include covering the PIN pad with a shield or with the other hand while typing. Although this protects PINs from visual attacks, acoustic emanations from the PIN pad itself open the door for another attack type. In this paper, we show the feasibility of an acoustic side-channel attack (called $\mathcal PinDrop$) to reconstruct PINs by profiling acoustic signatures of individual keys of a PIN pad. We demonstrate the practicality of $\mathcal PinDrop$ via two sets of data collection experiments involving two commercially available metal PIN pad models and 58 participants who entered a total of 5,800 5-digit PINs. We simulated two realistic attack scenarios: (1) a microphone placed near the ATM (0.3 m away) and (2) a real-time attacker (with a microphone) standing in the queue at a common courtesy distance of 2 m. In the former case, we show that $\mathcal PinDrop$ recovers 96% of 4-digit, and up to 94% of 5-digits, PINs. Whereas, at 2 m away, it recovers up to 57% of 4-digit, and up to 39% of 5-digit PINs in three attempts. We believe that these results are both significant and worrisome.

Appendix

1.1 A.1 8.1 Validation Results

Table 2 reports the results on the validation set for four different ML models. Results show that LR and SVC obtain the best results on PAD-1 and PAD-2, respectively.

1.2 B.2 8.2 Additional Results

In Fig. 8, we report the key accuracy results for PAD-2 (from both 0.3 m and 2 m). The results refer to the SVC model that achieved better performances on PAD-2.

In Fig. 9, we report an example for the digit “3” for all the four scenarios. All the other keys show similar behavior, highlighting no significant inter-class differences. Interestingly, we note a different distribution of classification errors between PAD-1 and PAD-2. In the first case, the error is uniformly distributed over all digits, in the second case, a higher concentration of errors is prominent around the true digit (i.e., digits 2, 5, and 6).

Table 2. PIN accuracies on the validation set for the investigated classifiers. The training set includes samples from five distinct attackers. The results show that for PAD-1 the best performing model is the Logistic Regression (LR), while for PAD-2 the best model is the SVC.

Fig. 8.

Key accuracy on the testing set for the best classifiers.

Fig. 9.

Digit “3” prediction heat maps for the four considered attack scenarios (the PIN pad layout is reported in Fig. 3). We reported the results for the experiment with 5 attackers and 500 digits entered per attacker

Figure 10 reports the PIN inference results within 3 attempts for PAD-2 and SVC model.

Fig. 10.

5-digit PINs inference performance within 3 attempts for the best classifiers

Figure 11 shows the results of $\mathcal PinDrop$ trained on the perturbed PAD-2 dataset (configuration 500 digits per attacker) in inferring 5-digit PINs within three attempts. The graphs report results similar to those obtained on PAD-1.

Fig. 11.

Impact of noise source and SNR in the inference of 5-digit PINs within three attempts for PAD-2 and 500 digits per attacker