arXiv subjects

Vitaly Shmatikov

Publications and source records attributed to Vitaly Shmatikov.

49 records · Page 3Linked to original sources

arXiv2019

Differential Privacy Has Disparate Impact on Model Accuracy

Differential privacy (DP) is a popular mechanism for training machine learning models with bounded leakage about the presence of specific points in the training data. The cost of differential privacy is a reduction in the model's accuracy. We demonstrate that in the neural networks trained using differentially private stochastic gradient descent (DP-SGD), this cost is not borne equally: accuracy of DP models drops much more for the underrepresented classes and subgroups. For example, a gender classification model trained using DP-SGD exhibits much lower accuracy for black faces than for white faces. Critically, this gap is bigger in the DP model than in the non-DP model, i.e., if the original model is unfair, the unfairness becomes worse once DP is applied. We demonstrate this effect for a variety of tasks and models, including sentiment analysis of text and image classification. We then explain why DP training mechanisms such as gradient clipping and noise addition have disproportionate effect on the underrepresented and more complex subgroups, resulting in a disparate reduction of model accuracy.

cs.LG↗

arXiv2019

How To Backdoor Federated Learning

Federated learning enables thousands of participants to construct a deep learning model without sharing their private training data with each other. For example, multiple smartphones can jointly train a next-word predictor for keyboards without revealing what individual users type. We demonstrate that any participant in federated learning can introduce hidden backdoor functionality into the joint global model, e.g., to ensure that an image classifier assigns an attacker-chosen label to images with certain features, or that a word predictor completes certain sentences with an attacker-chosen word. We design and evaluate a new model-poisoning methodology based on model replacement. An attacker selected in a single round of federated learning can cause the global model to immediately reach 100% accuracy on the backdoor task. We evaluate the attack under different assumptions for the standard federated-learning tasks and show that it greatly outperforms data poisoning. Our generic constrain-and-scale technique also evades anomaly detection-based defenses by incorporating the evasion into the attacker's loss function during training.

cs.CR↗

arXiv2019

Auditing Data Provenance in Text-Generation Models

To help enforce data-protection regulations such as GDPR and detect unauthorized uses of personal data, we develop a new \emph{model auditing} technique that helps users check if their data was used to train a machine learning model. We focus on auditing deep-learning models that generate natural-language text, including word prediction and dialog generation. These models are at the core of popular online services and are often trained on personal data such as users' messages, searches, chats, and comments. We design and evaluate a black-box auditing method that can detect, with very few queries to a model, if a particular user's texts were used to train it (among thousands of other users). We empirically show that our method can successfully audit well-generalized models that are not overfitted to the training data. We also analyze how text-generation models memorize word sequences and explain why this memorization makes them amenable to auditing.

cs.CR↗

arXiv2018

Exploiting Unintended Feature Leakage in Collaborative Learning

Collaborative machine learning and related techniques such as federated learning allow multiple participants, each with his own training dataset, to build a joint model by training locally and periodically exchanging model updates. We demonstrate that these updates leak unintended information about participants' training data and develop passive and active inference attacks to exploit this leakage. First, we show that an adversarial participant can infer the presence of exact data points -- for example, specific locations -- in others' training data (i.e., membership inference). Then, we show how this adversary can infer properties that hold only for a subset of the training data and are independent of the properties that the joint model aims to capture. For example, he can infer when a specific person first appears in the photos used to train a binary gender classifier. We evaluate our attacks on a variety of tasks, datasets, and learning configurations, analyze their limitations, and discuss possible defenses.

cs.CR↗

arXiv2018

Chiron: Privacy-preserving Machine Learning as a Service

Major cloud operators offer machine learning (ML) as a service, enabling customers who have the data but not ML expertise or infrastructure to train predictive models on this data. Existing ML-as-a-service platforms require users to reveal all training data to the service operator. We design, implement, and evaluate Chiron, a system for privacy-preserving machine learning as a service. First, Chiron conceals the training data from the service operator. Second, in keeping with how many existing ML-as-a-service platforms work, Chiron reveals neither the training algorithm nor the model structure to the user, providing only black-box access to the trained model. Chiron is implemented using SGX enclaves, but SGX alone does not achieve the dual goals of data privacy and model confidentiality. Chiron runs the standard ML training toolchain (including the popular Theano framework and C compiler) in an enclave, but the untrusted model-creation code from the service operator is further confined in a Ryoan sandbox to prevent it from leaking the training data outside the enclave. To support distributed training, Chiron executes multiple concurrent enclaves that exchange model parameters via a parameter server. We evaluate Chiron on popular deep learning models, focusing on benchmark image classification tasks such as CIFAR and ImageNet, and show that its training performance and accuracy of the resulting models are practical for common uses of ML-as-a-service.

cs.CR↗

arXiv2018

Fooling OCR Systems with Adversarial Text Images

We demonstrate that state-of-the-art optical character recognition (OCR) based on deep learning is vulnerable to adversarial images. Minor modifications to images of printed text, which do not change the meaning of the text to a human reader, cause the OCR system to "recognize" a different text where certain words chosen by the adversary are replaced by their semantic opposites. This completely changes the meaning of the output produced by the OCR system and by the NLP applications that use OCR for preprocessing their inputs.

cs.LG↗

arXiv2017

Machine Learning Models that Remember Too Much

Machine learning (ML) is becoming a commodity. Numerous ML frameworks and services are available to data holders who are not ML experts but want to train predictive models on their data. It is important that ML models trained on sensitive inputs (e.g., personal images or documents) not leak too much information about the training data. We consider a malicious ML provider who supplies model-training code to the data holder, does not observe the training, but then obtains white- or black-box access to the resulting model. In this setting, we design and implement practical algorithms, some of them very similar to standard ML techniques such as regularization and data augmentation, that "memorize" information about the training dataset in the model yet the model is as accurate and predictive as a conventionally trained model. We then explain how the adversary can extract memorized information from the model. We evaluate our techniques on standard ML tasks for image classification (CIFAR10), face recognition (LFW and FaceScrub), and text analysis (20 Newsgroups and IMDB). In all cases, we show how our algorithms create models that have high predictive power yet allow accurate extraction of subsets of their training data.

cs.CR↗

arXiv2017

Membership Inference Attacks against Machine Learning Models

We quantitatively investigate how machine learning models leak information about the individual data records on which they were trained. We focus on the basic membership inference attack: given a data record and black-box access to a model, determine if the record was in the model's training dataset. To perform membership inference against a target model, we make adversarial use of machine learning and train our own inference model to recognize differences in the target model's predictions on the inputs that it trained on versus the inputs that it did not train on. We empirically evaluate our inference techniques on classification models trained by commercial "machine learning as a service" providers such as Google and Amazon. Using realistic datasets and classification tasks, including a hospital discharge dataset whose membership is sensitive from the privacy perspective, we show that these models can be vulnerable to membership inference attacks. We then investigate the factors that influence this leakage and evaluate mitigation strategies.

cs.CR↗

arXiv2016

Defeating Image Obfuscation with Deep Learning

We demonstrate that modern image recognition methods based on artificial neural networks can recover hidden information from images protected by various forms of obfuscation. The obfuscation techniques considered in this paper are mosaicing (also known as pixelation), blurring (as used by YouTube), and P3, a recently proposed system for privacy-preserving photo sharing that encrypts the significant JPEG coefficients to make images unrecognizable by humans. We empirically show how to train artificial neural networks to successfully identify faces and recognize objects and handwritten digits even if the images are protected using any of the above obfuscation techniques.

cs.CR↗

arXiv2016

Towards a Privacy Research Roadmap for the Computing Community

Great advances in computing and communication technology are bringing many benefits to society, with transformative changes and financial opportunities being created in health care, transportation, education, law enforcement, national security, commerce, and social interactions. Many of these benefits, however, involve the use of sensitive personal data, and thereby raise concerns about privacy. Failure to address these concerns can lead to a loss of trust in the private and public institutions that handle personal data, and can stifle the independent thought and expression that is needed for our democracy to flourish. This report, sponsored by the Computing Community Consortium (CCC), suggests a roadmap for privacy research over the next decade, aimed at enabling society to appropriately control threats to privacy while enjoying the benefits of information technology and data science. We hope that it will be useful to the agencies of the Federal Networking and Information Technology Research and Development (NITRD) Program as they develop a joint National Privacy Research Strategy over the coming months. The report synthesizes input drawn from the privacy and computing communities submitted to both the CCC and NITRD, as well as past reports on the topic.

cs.CY↗

arXiv2016

Gone in Six Characters: Short URLs Considered Harmful for Cloud Services

Modern cloud services are designed to encourage and support collaboration. To help users share links to online documents, maps, etc., several services, including cloud storage providers such as Microsoft OneDrive and mapping services such as Google Maps, directly integrate URL shorteners that convert long, unwieldy URLs into short URLs, consisting of a domain such as 1drv.ms or goo.gl and a short token. In this paper, we demonstrate that the space of 5- and 6-character tokens included in short URLs is so small that it can be scanned using brute-force search. Therefore, all online resources that were intended to be shared with a few trusted friends or collaborators are effectively public and can be accessed by anyone. This leads to serious security and privacy vulnerabilities. In the case of cloud storage, we focus on Microsoft OneDrive. We show how to use short-URL enumeration to discover and read shared content stored in the OneDrive cloud, including even files for which the user did not generate a short URL. 7% of the OneDrive accounts exposed in this fashion allow anyone to write into them. Since cloud-stored files are automatically copied into users' personal computers and devices, this is a vector for large-scale, automated malware injection. In the case of online maps, we show how short-URL enumeration reveals the directions that users shared with each other. For many individual users, this enables inference of their residential addresses, true identities, and extremely sensitive locations they visited that, if publicly revealed, would violate medical and financial privacy.

cs.CR↗

arXiv2009

De-anonymizing Social Networks

Operators of online social networks are increasingly sharing potentially sensitive information about users and their relationships with advertisers, application developers, and data-mining researchers. Privacy is typically protected by anonymization, i.e., removing names, addresses, etc. We present a framework for analyzing privacy and anonymity in social networks and develop a new re-identification algorithm targeting anonymized social-network graphs. To demonstrate its effectiveness on real-world networks, we show that a third of the users who can be verified to have accounts on both Twitter, a popular microblogging service, and Flickr, an online photo-sharing site, can be re-identified in the anonymous Twitter graph with only a 12% error rate. Our de-anonymization algorithm is based purely on the network topology, does not require creation of a large number of dummy "sybil" nodes, is robust to noise and all existing defenses, and works even when the overlap between the target network and the adversary's auxiliary information is small.

cs.CR↗

arXiv2007

How To Break Anonymity of the Netflix Prize Dataset

We present a new class of statistical de-anonymization attacks against high-dimensional micro-data, such as individual preferences, recommendations, transaction records and so on. Our techniques are robust to perturbation in the data and tolerate some mistakes in the adversary's background knowledge. We apply our de-anonymization methodology to the Netflix Prize dataset, which contains anonymous movie ratings of 500,000 subscribers of Netflix, the world's largest online movie rental service. We demonstrate that an adversary who knows only a little bit about an individual subscriber can easily identify this subscriber's record in the dataset. Using the Internet Movie Database as the source of background knowledge, we successfully identified the Netflix records of known users, uncovering their apparent political preferences and other potentially sensitive information.

cs.CR↗