Hubness is a phenomenon in high-dimensional vector spaces where a point from the natural distribution is unusually close to many other points. This is a well-known problem in information retrieval that causes some items to accidentally (and incorrectly) appear relevant to many queries. In this paper, we investigate how attackers can exploit hubness to turn any image or audio input in a multi-modal retrieval system into an adversarial hub. Adversarial hubs can be used to inject universal adversarial content (e.g., spam) that will be retrieved in response to thousands of different queries, and also for targeted attacks on queries related to specific, attacker-chosen concepts. We present a method for creating adversarial hubs and evaluate the resulting hubs on benchmark multi-modal retrieval datasets and an image-to-image retrieval system implemented by Pinecone, a popular vector database. For example, in text-caption-to-image retrieval, a single adversarial hub, generated using 100 random queries, is retrieved as the top-1 most relevant image for more than 21,000 out of 25,000 test queries (by contrast, the most common natural hub is the top-1 response to only 102 queries), demonstrating the strong generalization capabilities of adversarial hubs. We also investigate whether techniques for mitigating natural hubness can also mitigate adversarial hubs, and show that they are not effective against hubs that target queries related to specific concepts.

Transfer learning is a popular method for tuning pretrained (upstream) models for different downstream tasks using limited data and computational resources. We study how an adversary with control over an upstream model used in transfer learning can conduct property inference attacks on a victim's tuned downstream model. For example, to infer the presence of images of a specific individual in the downstream training set. We demonstrate attacks in which an adversary can manipulate the upstream model to conduct highly effective and specific property inference attacks (AUC score > 0.9), without incurring significant performance loss on the main task. The main idea of the manipulation is to make the upstream model generate activations (intermediate features) with different distributions for samples with and without a target property, thus enabling the adversary to distinguish easily between downstream models trained with and without training examples that have the target property. Our code is available at https://github.com/yulongt23/Transfer-Inference.

We study indiscriminate poisoning for linear learners where an adversary injects a few crafted examples into the training data with the goal of forcing the induced model to incur higher test error. Inspired by the observation that linear learners on some datasets are able to resist the best known attacks even without any defenses, we further investigate whether datasets can be inherently robust to indiscriminate poisoning attacks for linear learners. For theoretical Gaussian distributions, we rigorously characterize the behavior of an optimal poisoning attack, defined as the poisoning strategy that attains the maximum risk of the induced model at a given poisoning budget. Our results prove that linear learners can indeed be robust to indiscriminate poisoning if the class-wise data distributions are well-separated with low variance and the size of the constraint set containing all permissible poisoning points is also small. These findings largely explain the drastic variation in empirical attack performance of the state-of-the-art poisoning attacks on linear learners across benchmark datasets, making an important initial step towards understanding the underlying reasons some learning tasks are vulnerable to data poisoning attacks.

Model compression is a widely-used approach for reducing the size of deep learning models without much accuracy loss, enabling resource-hungry models to be compressed for use on resource-constrained devices. In this paper, we study the risk that model compression could provide an opportunity for adversaries to inject stealthy backdoors. In a backdoor attack on a machine learning model, an adversary produces a model that performs well on normal inputs but outputs targeted misclassifications on inputs containing a small trigger pattern. We design stealthy backdoor attacks such that the full-sized model released by adversaries appears to be free from backdoors (even when tested using state-of-the-art techniques), but when the model is compressed it exhibits a highly effective backdoor. We show this can be done for two common model compression techniques—model pruning and model quantization—even in settings where the adversary has limited knowledge of how the particular compression will be done. Our findings demonstrate the importance of performing security tests on the models that will actually be deployed not in their precompressed version. Our implementation is available athttps://github.com/yulongtzzz/Stealthy-Backdoors-as-Compression-Artifacts.

In a poisoning attack, an adversary who controls a small fraction of the training data attempts to select that data, so a model is induced that misbehaves in a particular way. We consider poisoning attacks against convex machine learning models and propose an efficient poisoning attack designed to induce a model specified by the adversary. Unlike previous model-targeted poisoning attacks, our attack comes with provable convergence to any attainable target model. We also provide a lower bound on the minimum number of poisoning points needed to achieve a given target model. Our method uses online convex optimization and finds poisoning points incrementally. This provides more flexibility than previous attacks which require an a priori assumption about the number of poisoning points. Our attack is the first model-targeted poisoning attack that provides provable convergence for convex models. In our experiments, it either exceeds or matches state-of-the-art attacks in terms of attack success rate and distance to the target model.

Mobile apps include privacy settings that allow their users to configure how their data should be shared. These settings, however, are often hard to locate and hard to understand by the users, even in popular apps, such as Facebook. More seriously, they are often set to share user data by default, exposing her privacy without proper consent. In this paper, we report the first systematic study on the problem, which is made possible through an in-depth analysis of user perception of the privacy settings. More specifically, we first conduct two user studies (involving nearly one thousand users) to understand privacy settings from the user's perspective, and identify these hard-to-find settings. Then we select 14 features that uniquely characterize such hidden privacy settings and utilize a novel technique called semantics- based UI tracing to extract them from a given app. On top of these features, a classifier is trained to automatically discover the hidden privacy settings, which together with other innovations, has been implemented into a tool called Hound. Over our labeled data set, the tool achieves an accuracy of 93.54%. Further running it on 100,000 latest apps from both Google Play and third-party markets, we find that over a third (36.29%) of the privacy settings identified from these apps are “hidden”. Looking into these settings, we observe that they become hard to discover and hard to understand primarily due to the problematic categorization on the apps' user interfaces and/or confusing descriptions. Further importantly, though more privacy options have been offered to the user over time, also discovered is the persistence of their usability issue, which becomes even more serious, e.g., originally easy-to-find settings now harder to locate. And among all such hidden privacy settings, 82.16% are set to leak user privacy by default. We provide suggestions for improving the usability of these privacy settings at the end of our study.