David Evans

Hojjat Aghakhani, Wei Dai, Andre Manoel et al. · microsoft-research, mit

With tools like GitHub Copilot, automatic code suggestion is no longer a dream in software engineering. These tools, based on large language models, are typically trained on massive corpora of code mined from unvetted public sources. As a result, these models are susceptible to data poisoning attacks where an adversary manipulates the model's training by injecting malicious data. Poisoning attacks could be designed to influence the model's suggestions at run time for chosen contexts, such as inducing the model into suggesting insecure code payloads. To achieve this, prior attacks explicitly inject the insecure code payload into the training data, making the poison data detectable by static analysis tools that can remove such malicious data from the training set. In this work, we demonstrate two novel attacks, COVERT and TROJANPUZZLE, that can bypass static analysis by planting malicious poison data in out-of-context regions such as docstrings. Our most novel attack, TROJANPUZZLE, goes one step further in generating less suspicious poison data by never explicitly including certain (suspicious) parts of the payload in the poison data, while still inducing a model that suggests the entire payload when completing code (i.e., outside docstrings). This makes TROJANPUZZLE robust against signature-based dataset-cleansing methods that can filter out suspicious sequences from the training data. Our evaluation against models of two sizes demonstrates that both COVERT and TROJANPUZZLE have significant implications for practitioners when selecting code used to train or tune code-suggestion models.

Fatemehsadat Mireshghallah, Archit Uniyal, Tianhao Wang et al. · mit

Large language models are shown to present privacy risks through memorization of training data, and several recent works have studied such risks for the pre-training phase. Little attention, however, has been given to the fine-tuning phase and it is not well understood how different fine-tuning methods (such as fine-tuning the full model, the model head, and adapter) compare in terms of memorization risk. This presents increasing concern as the "pre-train and fine-tune" paradigm proliferates. In this paper, we empirically study memorization of fine-tuning methods using membership inference and extraction attacks, and show that their susceptibility to attacks is very different. We observe that fine-tuning the head of the model has the highest susceptibility to attacks, whereas fine-tuning smaller adapters appears to be less vulnerable to known extraction attacks.

Bargav Jayaraman, David Evans

Models can expose sensitive information about their training data. In an attribute inference attack, an adversary has partial knowledge of some training records and access to a model trained on those records, and infers the unknown values of a sensitive feature of those records. We study a fine-grained variant of attribute inference we call \emph{sensitive value inference}, where the adversary's goal is to identify with high confidence some records from a candidate set where the unknown attribute has a particular sensitive value. We explicitly compare attribute inference with data imputation that captures the training distribution statistics, under various assumptions about the training data available to the adversary. Our main conclusions are: (1) previous attribute inference methods do not reveal more about the training data from the model than can be inferred by an adversary without access to the trained model, but with the same knowledge of the underlying distribution as needed to train the attribute inference attack; (2) black-box attribute inference attacks rarely learn anything that cannot be learned without the model; but (3) white-box attacks, which we introduce and evaluate in the paper, can reliably identify some records with the sensitive value attribute that would not be predicted without having access to the model. Furthermore, we show that proposed defenses such as differentially private training and removing vulnerable records from training do not mitigate this privacy risk. The code for our experiments is available at \url{https://github.com/bargavj/EvaluatingDPML}.

Yulong Tian, Fnu Suya, Anshuman Suri et al.

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.

Anshuman Suri, Yifu Lu, Yanjin Chen et al.

A distribution inference attack aims to infer statistical properties of data used to train machine learning models. These attacks are sometimes surprisingly potent, but the factors that impact distribution inference risk are not well understood and demonstrated attacks often rely on strong and unrealistic assumptions such as full knowledge of training environments even in supposedly black-box threat scenarios. To improve understanding of distribution inference risks, we develop a new black-box attack that even outperforms the best known white-box attack in most settings. Using this new attack, we evaluate distribution inference risk while relaxing a variety of assumptions about the adversary's knowledge under black-box access, like known model architectures and label-only access. Finally, we evaluate the effectiveness of previously proposed defenses and introduce new defenses. We find that although noise-based defenses appear to be ineffective, a simple re-sampling defense can be highly effective. Code is available at https://github.com/iamgroot42/dissecting_distribution_inference

Josephine Lamp, Mark Derdzinski, Christopher Hannemann et al.

We focus on the problem of generating high-quality, private synthetic glucose traces, a task generalizable to many other time series sources. Existing methods for time series data synthesis, such as those using Generative Adversarial Networks (GANs), are not able to capture the innate characteristics of glucose data and cannot provide any formal privacy guarantees without severely degrading the utility of the synthetic data. In this paper we present GlucoSynth, a novel privacy-preserving GAN framework to generate synthetic glucose traces. The core intuition behind our approach is to conserve relationships amongst motifs (glucose events) within the traces, in addition to temporal dynamics. Our framework incorporates differential privacy mechanisms to provide strong formal privacy guarantees. We provide a comprehensive evaluation on the real-world utility of the data using 1.2 million glucose traces; GlucoSynth outperforms all previous methods in its ability to generate high-quality synthetic glucose traces with strong privacy guarantees.

Ahmed Salem, Giovanni Cherubin, David Evans et al.

Deploying machine learning models in production may allow adversaries to infer sensitive information about training data. There is a vast literature analyzing different types of inference risks, ranging from membership inference to reconstruction attacks. Inspired by the success of games (i.e., probabilistic experiments) to study security properties in cryptography, some authors describe privacy inference risks in machine learning using a similar game-based style. However, adversary capabilities and goals are often stated in subtly different ways from one presentation to the other, which makes it hard to relate and compose results. In this paper, we present a game-based framework to systematize the body of knowledge on privacy inference risks in machine learning. We use this framework to (1) provide a unifying structure for definitions of inference risks, (2) formally establish known relations among definitions, and (3) to uncover hitherto unknown relations that would have been difficult to spot otherwise.

Bargav Jayaraman, Esha Ghosh, Melissa Chase et al.

Pre-trained large language models, such as GPT\nobreakdash-2 and BERT, are often fine-tuned to achieve state-of-the-art performance on a downstream task. One natural example is the ``Smart Reply'' application where a pre-trained model is tuned to provide suggested responses for a given query message. Since the tuning data is often sensitive data such as emails or chat transcripts, it is important to understand and mitigate the risk that the model leaks its tuning data. We investigate potential information leakage vulnerabilities in a typical Smart Reply pipeline. We consider a realistic setting where the adversary can only interact with the underlying model through a front-end interface that constrains what types of queries can be sent to the model. Previous attacks do not work in these settings, but require the ability to send unconstrained queries directly to the model. Even when there are no constraints on the queries, previous attacks typically require thousands, or even millions, of queries to extract useful information, while our attacks can extract sensitive data in just a handful of queries. We introduce a new type of active extraction attack that exploits canonical patterns in text containing sensitive data. We show experimentally that it is possible for an adversary to extract sensitive user information present in the training data, even in realistic settings where all interactions with the model must go through a front-end that limits the types of queries. We explore potential mitigation strategies and demonstrate empirically how differential privacy appears to be a reasonably effective defense mechanism to such pattern extraction attacks.

Valentin Hartmann, Anshuman Suri, Vincent Bindschaedler et al.

Large Language Models (LLMs) are advancing at a remarkable pace, with myriad applications under development. Unlike most earlier machine learning models, they are no longer built for one specific application but are designed to excel in a wide range of tasks. A major part of this success is due to their huge training datasets and the unprecedented number of model parameters, which allow them to memorize large amounts of information contained in the training data. This memorization goes beyond mere language, and encompasses information only present in a few documents. This is often desirable since it is necessary for performing tasks such as question answering, and therefore an important part of learning, but also brings a whole array of issues, from privacy and security to copyright and beyond. LLMs can memorize short secrets in the training data, but can also memorize concepts like facts or writing styles that can be expressed in text in many different ways. We propose a taxonomy for memorization in LLMs that covers verbatim text, facts, ideas and algorithms, writing styles, distributional properties, and alignment goals. We describe the implications of each type of memorization - both positive and negative - for model performance, privacy, security and confidentiality, copyright, and auditing, and ways to detect and prevent memorization. We further highlight the challenges that arise from the predominant way of defining memorization with respect to model behavior instead of model weights, due to LLM-specific phenomena such as reasoning capabilities or differences between decoding algorithms. Throughout the paper, we describe potential risks and opportunities arising from memorization in LLMs that we hope will motivate new research directions.

Hannah Chen, Yangfeng Ji, David Evans

Traditional (fickle) adversarial examples involve finding a small perturbation that does not change an input's true label but confuses the classifier into outputting a different prediction. Conversely, obstinate adversarial examples occur when an adversary finds a small perturbation that preserves the classifier's prediction but changes the true label of an input. Adversarial training and certified robust training have shown some effectiveness in improving the robustness of machine learnt models to fickle adversarial examples. We show that standard adversarial training methods focused on reducing vulnerability to fickle adversarial examples may make a model more vulnerable to obstinate adversarial examples, with experiments for both natural language inference and paraphrase identification tasks. To counter this phenomenon, we introduce Balanced Adversarial Training, which incorporates contrastive learning to increase robustness against both fickle and obstinate adversarial examples.

Evan Rose, Fnu Suya, David Evans

Machine learning is susceptible to poisoning attacks, in which an attacker controls a small fraction of the training data and chooses that data with the goal of inducing some behavior unintended by the model developer in the trained model. We consider a realistic setting in which the adversary with the ability to insert a limited number of data points attempts to control the model's behavior on a specific subpopulation. Inspired by previous observations on disparate effectiveness of random label-flipping attacks on different subpopulations, we investigate the properties that can impact the effectiveness of state-of-the-art poisoning attacks against different subpopulations. For a family of 2-dimensional synthetic datasets, we empirically find that dataset separability plays a dominant role in subpopulation vulnerability for less separable datasets. However, well-separated datasets exhibit more dependence on individual subpopulation properties. We further discover that a crucial subpopulation property is captured by the difference in loss on the clean dataset between the clean model and a target model that misclassifies the subpopulation, and a subpopulation is much easier to attack if the loss difference is small. This property also generalizes to high-dimensional benchmark datasets. For the Adult benchmark dataset, we show that we can find semantically-meaningful subpopulation properties that are related to the susceptibilities of a selected group of subpopulations. The results in this paper are accompanied by a fully interactive web-based visualization of subpopulation poisoning attacks found at https://uvasrg.github.io/visualizing-poisoning

Anshuman Suri, David Evans

Distribution inference, sometimes called property inference, infers statistical properties about a training set from access to a model trained on that data. Distribution inference attacks can pose serious risks when models are trained on private data, but are difficult to distinguish from the intrinsic purpose of statistical machine learning -- namely, to produce models that capture statistical properties about a distribution. Motivated by Yeom et al.'s membership inference framework, we propose a formal definition of distribution inference attacks that is general enough to describe a broad class of attacks distinguishing between possible training distributions. We show how our definition captures previous ratio-based property inference attacks as well as new kinds of attack including revealing the average node degree or clustering coefficient of a training graph. To understand distribution inference risks, we introduce a metric that quantifies observed leakage by relating it to the leakage that would occur if samples from the training distribution were provided directly to the adversary. We report on a series of experiments across a range of different distributions using both novel black-box attacks and improved versions of the state-of-the-art white-box attacks. Our results show that inexpensive attacks are often as effective as expensive meta-classifier attacks, and that there are surprising asymmetries in the effectiveness of attacks. Code is available at https://github.com/iamgroot42/FormEstDistRisks

Yulong Tian, Fnu Suya, Fengyuan Xu et al.

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. 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. 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 highly effective backdoors. We show this can be done for two common model compression techniques -- model pruning and model quantization. Our findings demonstrate how an adversary may be able to hide a backdoor as a compression artifact, and show the importance of performing security tests on the models that will actually be deployed not their precompressed version.

Hannah Li, David Evans

Vulnerabilities in password managers are unremitting because current designs provide large attack surfaces, both at the client and server. We describe and evaluate Horcrux, a password manager that is designed holistically to minimize and decentralize trust, while retaining the usability of a traditional password manager. The prototype Horcrux client, implemented as a Firefox add-on, is split into two components, with code that has access to the user's master's password and any key material isolated into a small auditable component, separate from the complexity of managing the user interface. Instead of exposing actual credentials to the DOM, a dummy username and password are autofilled by the untrusted component. The trusted component intercepts and modifies POST requests before they are encrypted and sent over the network. To avoid trusting a centralized store, stored credentials are secret-shared over multiple servers. To provide domain and username privacy, while maintaining resilience to off-line attacks on a compromised password store, we incorporate cuckoo hashing in a way that ensures an attacker cannot determine if a guessed master password is correct. Our approach only works for websites that do not manipulate entered credentials in the browser client, so we conducted a large-scale experiment that found the technique appears to be compatible with over 98% of tested login forms.