Neural networks (NNs) have become one of the most important tools for artificial intelligence. Well-designed and trained NNs can perform inference (e.g., make decisions or predictions) on unseen inputs with high accuracy. Using NNs often involves sensitive data: Depending on the specific use case, the input to the NN and/or the internals of the NN (e.g., the weights and biases) may be sensitive. Thus, there is a need for techniques for performing NN inference securely, ensuring that sensitive data remain secret. In the past few years, several approaches have been proposed for secure neural network inference. These approaches achieve better and better results in terms of efficiency, security, accuracy, and applicability, thus making big progress toward practical secure neural network inference. The proposed approaches make use of many different techniques, such as homomorphic encryption and secure multi-party computation. The aim of this article is to give an overview of the main approaches proposed so far, their different properties, and the techniques used. In addition, remaining challenges toward large-scale deployments are identified.
Contact discovery allows users of mobile messengers to conveniently connect with people in their address book. In this work, we demonstrate that severe privacy issues exist in currently deployed contact discovery methods and propose suitable mitigations. Our study of three popular messengers (WhatsApp, Signal, and Telegram) shows that large-scale crawling attacks are (still) possible. Using an accurate database of mobile phone number prefixes and very few resources, we queried 10 % of US mobile phone numbers for WhatsApp and 100 % for Signal. For Telegram, we find that its API exposes a wide range of sensitive information, even about numbers not registered with the service. We present interesting (cross-messenger) usage statistics, which also reveal that very few users change the default privacy settings. Furthermore, we demonstrate that currently deployed hashing-based contact discovery protocols are severely broken by comparing three methods for efficient hash reversal. Most notably, we show that with the password cracking tool “JTR,” we can iterate through the entire worldwide mobile phone number space in < 150 s on a consumer-grade GPU. We also propose a significantly improved rainbow table construction for non-uniformly distributed input domains that is of independent interest. Regarding mitigations, we most notably propose two novel rate-limiting schemes: our incremental contact discovery for services without server-side contact storage strictly improves over Signal’s current approach while being compatible with private set intersection, whereas our differential scheme allows even stricter rate limits at the overhead for service providers to store a small constant-size state that does not reveal any contact information.
We show that perceptual hashing, crucial for detecting and removing image-based sexual abuse (IBSA) online, faces vulnerabilities from low-budget inversion attacks based on generative AI. This jeopardizes the privacy of users, especially vulnerable groups. We advocate to implement secure hash matching in IBSA removal tools to mitigate potentially fatal consequences.