Deep learning models are vulnerable to privacy attacks due to their tendency to memorize individual training examples. Theoretically-sound defenses such as differential privacy can defend against this threat, but model performance often suffers. Empirical defenses may thwart existing attacks while maintaining model performance but do not offer any robust theoretical guarantees. In this paper, we explore a new strategy based on the concept of plausible deniability. We introduce a training algorithm called **P**lausibly **D**eniable **S**tochastic **G**radient **D**escent (PD-SGD). The core of this approach is a rejection sampling technique, which probabilistically prevents updating model parameters whenever a mini-batch cannot be plausibly denied. We provide theoretical results showing that PD-SGD effectively mitigates privacy leakage from individual data points. Experiments demonstrate the scalability of PD-SGD and the favorable privacy-utility trade-off it offers compared to existing defense methods.

Automatic Speech Recognition (ASR) systems convert speech into text and can be placed into two broad categories: traditional and fully end-to-end. Both types have been shown to be vulnerable to adversarial audio examples that sound benign to the human ear but force the ASR to produce malicious transcriptions. Of these attacks, only the “psychoacoustic” attacks can create examples with relatively imperceptible perturbations, as they leverage the knowledge of the human auditory system. Unfortunately, existing psychoacoustic attacks can only be applied against traditional models, and are obsolete against the newer, fully end-to-end ASRs. In this paper, we propose an equalization-based psychoacoustic attack that can exploit both traditional and fully end-to-end ASRs. We successfully demonstrate our attack against real-world ASRs that include DeepSpeech and Wav2Letter. Moreover, we employ a user study to verify that our method creates low audible distortion. Specifically, 80 of the 100 participants voted in favor of \textit{all} our attack audio samples as less noisier than the existing state-of-the-art attack. Through this, we demonstrate both types of existing ASR pipelines can be exploited with minimum degradation to attack audio quality.

Automatic speech recognition and voice identification systems are being deployed in a wide array of applications, from providing control mechanisms to devices lacking traditional interfaces, to the automatic transcription of conversations and authentication of users. Many of these applications have significant security and privacy considerations. We develop attacks that force mistranscription and misidentification in state of the art systems, with minimal impact on human comprehension. Processing pipelines for modern systems are comprised of signal preprocessing and feature extraction steps, whose output is fed to a machine-learned model. Prior work has focused on the models, using white-box knowledge to tailor model-specific attacks. We focus on the pipeline stages before the models, which (unlike the models) are quite similar across systems. As such, our attacks are black-box, transferable, can be tuned to require zero queries to the target, and demonstrably achieve mistranscription and misidentification rates as high as 100% by modifying only a few frames of audio. We perform a study via Amazon Mechanical Turk demonstrating that there is no statistically significant difference between human perception of regular and perturbed audio. Our findings suggest that models may learn aspects of speech that are generally not perceived by human subjects, but that are crucial for model accuracy.

Speech and speaker recognition systems are employed in a variety of applications, from personal assistants to telephony surveillance and biometric authentication. The wide deployment of these systems has been made possible by the improved accuracy in neural networks. Like other systems based on neural networks, recent research has demonstrated that speech and speaker recognition systems are vulnerable to attacks using manipulated inputs. However, as we demonstrate in this paper, the end-to-end architecture of speech and speaker systems and the nature of their inputs make attacks and defenses against them substantially different than those in the image space. We demonstrate this first by systematizing existing research in this space and providing a taxonomy through which the community can evaluate future work. We then demonstrate experimentally that attacks against these models almost universally fail to transfer. In so doing, we argue that substantial additional work is required to provide adequate mitigations in this space.

Voice controlled interfaces have vastly improved the usability of many devices (e.g., headless IoT systems). Unfortunately, the lack of authentication for these interfaces has also introduced command injection vulnerabilities - whether via compromised IoT devices, television ads or simply malicious nearby neighbors, causing such devices to perform unauthenticated sensitive commands is relatively easy. We address these weaknesses with Two Microphone Authentication (2MA), which takes advantage of the presence of multiple ambient and personal devices operating in the same area. We develop an embodiment of 2MA that combines approximate localization through Direction of Arrival (DOA) techniques with Robust Audio Hashes (RSHs). Our results show that our 2MA system can localize a source to within a narrow physical cone ($<30^\circ $) with zero false positives, eliminate replay attacks and prevent the injection of inaudible/hidden commands. As such, we dramatically increase the difficulty for an adversary to carry out such attacks and demonstrate that 2MA is an effective means of authenticating and localizing voice commands.

The global telephone network is relied upon by billions every day. Central to its operation is the Signaling System 7 (SS7) protocol, which is used for setting up calls, managing mobility, and facilitating many other network services. This protocol was originally built on the assumption that only a small number of trusted parties would be able to directly communicate with its core infrastructure. As a result, SS7 - as a feature - allows all parties with core access to redirect and intercept calls for any subscriber anywhere in the world. Unfortunately, increased interconnectivity with the SS7 network has led to a growing number of illicit call redirection attacks. We address such attacks with Sonar, a system that detects the presence of SS7 redirection attacks by securely measuring call audio round-trip times between telephony devices. This approach works because redirection attacks force calls to travel longer physical distances than usual, thereby creating longer end-to-end delay. We design and implement a distance bounding-inspired protocol that allows us to securely characterize the round-trip time between the two endpoints. We then use custom hardware deployed in 10 locations across the United States and a redirection testbed to characterize how distance affects round trip time in phone networks. We develop a model using this testbed and show Sonar is able to detect 70.9% of redirected calls between call endpoints of varying attacker proximity (300-7100 miles) with low false positive rates (0.3%). Finally, we ethically perform actual SS7 redirection attacks on our own devices with the help of an industry partner to demonstrate that Sonar detects 100% of such redirections in a real network (with no false positives). As such, we demonstrate that telephone users can reliably detect SS7 redirection attacks and protect the integrity of their calls.