Audio deepfakes represent a rising threat to trust in our daily communications. In response to this, the research community has developed a wide array of detection techniques aimed at preventing such attacks from deceiving users. Unfortunately, the creation of these defenses has generally overlooked the most important element of the system - the user themselves. As such, it is not clear whether current mechanisms augment, hinder, or simply contradict human classification of deepfakes. In this paper, we perform the first large-scale user study on deepfake detection. We recruit over 1,200 users and present them with samples from the three most widely-cited deepfake datasets. We then quantitatively compare performance and qualitatively conduct thematic analysis to motivate and understand the reasoning behind user decisions and differences from machine classifications. Our results show that users correctly classify human audio at significantly higher rates than machine learning models, and rely on linguistic features and intuition when performing classification. However, users are also regularly misled by pre-conceptions about the capabilities of generated audio (e.g., that accents and background sounds are indicative of humans). Finally, machine learning models suffer from significantly higher false positive rates, and experience false negatives that humans correctly classify when issues of quality or robotic characteristics are reported. By analyzing user behavior across multiple deepfake datasets, our study demonstrates the need to more tightly compare user and machine learning performance, and to target the latter towards areas where humans are less likely to successfully identify threats.

Audio Rogue Base Stations (ARBSs) allow an adversary to intercept cellular calls. These devices represent a substantial escalation in the threat posed by traditional rogue base stations, which only collect device identity information. This paper presents the first technique for detecting call eavesdropping via an ARBS. Our system, which we call LeopardSeal, uses distance bounding over the call audio channel to determine whether or not extra wireless hops (and therefore increased audio delay) that are characteristic of ARBSs are present during a call. We implement a proof of concept ARBS using open-source guides and perform a measurement study across the United States. We demonstrate the ability to detect all attacks (with zero false positives) due to a statistically significant difference in round trip times between benign and attack call audio (i.e., t-test: p ≪ 0.01) due to the large cost of additional wireless hops. Through this effort, we demonstrate the ability to robustly detect these eavesdropping devices.

Bluetooth is overwhelmingly the protocol of choice for personal area networking, and the Bluetooth Classic standard has been in continuous use for over 20 years. Bluetooth devices make themselves Discoverable to communicate, but best practice to protect privacy is to ensure that devices remain in Non-Discoverable mode. This paper demonstrates the futility of protecting devices by making them Non-Discoverable. We introduce the Blue’s Clues attack, which presents the first direct, non-disruptive approach to fully extracting the permanent, unique Bluetooth MAC identifier from targeted devices in Non-Discoverable mode. We also demonstrate that we can fully characterize device capabilities and retrieve identifiers, some of which we discover often contain identifying information about the device owner. We demonstrate Blue’s Clues using a software-defined radio and mounting the attack over the air against both our own devices and, with institutional approval, throughout a public building. We find that a wide variety of Bluetooth devices can be uniquely identified in less than 10 seconds on average, with affected devices ranging from smartphones and headphones to gas pump skimmers and nanny-cams, spanning all versions of the Bluetooth Classic standard. While we provide potential mitigation against attacks, Blue’s Clues forces a reassessment of over 20 years of best practices for protecting devices against discovery.