When building machine learning models using sensitive data, organizations should ensure that the data processed in such systems is adequately protected. For projects involving machine learning on personal data, Article 35 of the GDPR mandates it to perform a Data Protection Impact Assessment (DPIA). In addition to the threats of illegitimate access to data through security breaches, machine learning models pose an additional privacy risk to the data by indirectly revealing about it through the model predictions and parameters. Guidances released by the Information Commissioner's Office (UK) and the National Institute of Standards and Technology (US) emphasize on the threat to data from models and recommend organizations to account for and estimate these risks to comply with data protection regulations. Hence, there is an immediate need for a tool that can quantify the privacy risk to data from models. In this paper, we focus on this indirect leakage about training data from machine learning models. We present ML Privacy Meter, a tool that can quantify the privacy risk to data from models through state of the art membership inference attack techniques. We discuss how this tool can help practitioners in compliance with data protection regulations, when deploying machine learning models.
Federated learning algorithms are developed both for efficiency reasons and to ensure the privacy and confidentiality of personal and business data, respectively. Despite no data being shared explicitly, recent studies showed that the mechanism could still leak sensitive information. Hence, secure aggregation is utilized in many real-world scenarios to prevent attribution to specific participants. In this paper, we focus on the quality of individual training datasets and show that such quality information could be inferred and attributed to specific participants even when secure aggregation is applied. Specifically, through a series of image recognition experiments, we infer the relative quality ordering of participants. Moreover, we apply the inferred quality information to detect misbehaviours, to stabilize training performance, and to measure the individual contributions of participants.
Distributed machine learning generally aims at training a global model based on distributed data without collecting all the data to a centralized location, where two different approaches have been proposed: collecting and aggregating local models (federated learning) and collecting and training over representative data summaries (coreset). While each approach preserves data privacy to some extent thanks to not sharing the raw data, the exact extent of protection is unclear under sophisticated attacks that try to infer the raw data from the shared information. We present the first comparison between the two approaches in terms of target model accuracy, communication cost, and data privacy, where the last is measured by the accuracy of a state-of-the-art attack strategy called the membership inference attack. Our experiments quantify the accuracy-privacy-cost tradeoff of each approach, and reveal a nontrivial comparison that can be used to guide the design of model training processes.
Deep learning models often raise privacy concerns as they leak information about their training data. This enables an adversary to determine whether a data point was in a model's training set by conducting a membership inference attack (MIA). Prior work has conjectured that regularization techniques, which combat overfitting, may also mitigate the leakage. While many regularization mechanisms exist, their effectiveness against MIAs has not been studied systematically, and the resulting privacy properties are not well understood. We explore the lower bound for information leakage that practical attacks can achieve. First, we evaluate the effectiveness of 8 mechanisms in mitigating two recent MIAs, on three standard image classification tasks. We find that certain mechanisms, such as label smoothing, may inadvertently help MIAs. Second, we investigate the potential of improving the resilience to MIAs by combining complementary mechanisms. Finally, we quantify the opportunity of future MIAs to compromise privacy by designing a white-box `distance-to-confident' (DtC) metric, based on adversarial sample crafting. Our metric reveals that, even when existing MIAs fail, the training samples may remain distinguishable from test samples. This suggests that regularization mechanisms can provide a false sense of privacy, even when they appear effective against existing MIAs.
Recent studies propose membership inference (MI) attacks on deep models, where the goal is to infer if a sample has been used in the training process. Despite their apparent success, these studies only report accuracy, precision, and recall of the positive class (member class). Hence, the performance of these attacks have not been clearly reported on negative class (non-member class). In this paper, we show that the way the MI attack performance has been reported is often misleading because they suffer from high false positive rate or false alarm rate (FAR) that has not been reported. FAR shows how often the attack model mislabel non-training samples (non-member) as training (member) ones. The high FAR makes MI attacks fundamentally impractical, which is particularly more significant for tasks such as membership inference where the majority of samples in reality belong to the negative (non-training) class. Moreover, we show that the current MI attack models can only identify the membership of misclassified samples with mediocre accuracy at best, which only constitute a very small portion of training samples. We analyze several new features that have not been comprehensively explored for membership inference before, including distance to the decision boundary and gradient norms, and conclude that deep models' responses are mostly similar among train and non-train samples. We conduct several experiments on image classification tasks, including MNIST, CIFAR-10, CIFAR-100, and ImageNet, using various model architecture, including LeNet, AlexNet, ResNet, etc. We show that the current state-of-the-art MI attacks cannot achieve high accuracy and low FAR at the same time, even when the attacker is given several advantages. The source code is available at https://github.com/shrezaei/MI-Attack.
We study membership inference in settings where some of the assumptions typically used in previous research are relaxed. First, we consider skewed priors, to cover cases such as when only a small fraction of the candidate pool targeted by the adversary are actually members and develop a PPV-based metric suitable for this setting. This setting is more realistic than the balanced prior setting typically considered by researchers. Second, we consider adversaries that select inference thresholds according to their attack goals and develop a threshold selection procedure that improves inference attacks. Since previous inference attacks fail in imbalanced prior setting, we develop a new inference attack based on the intuition that inputs corresponding to training set members will be near a local minimum in the loss function, and show that an attack that combines this with thresholds on the per-instance loss can achieve high PPV even in settings where other attacks appear to be ineffective. Code for our experiments can be found here: https://github.com/bargavj/EvaluatingDPML.
Deep Learning (DL) techniques allow ones to train models from a dataset to solve tasks. DL has attracted much interest given its fancy performance and potential market value, while security issues are amongst the most colossal concerns. However, the DL models may be prone to the membership inference attack, where an attacker determines whether a given sample is from the training dataset. Efforts have been made to hinder the attack but unfortunately, they may lead to a major overhead or impaired usability. In this paper, we propose and implement DAMIA, leveraging Domain Adaptation (DA) as a defense aginist membership inference attacks. Our observation is that during the training process, DA obfuscates the dataset to be protected using another related dataset, and derives a model that underlyingly extracts the features from both datasets. Seeing that the model is obfuscated, membership inference fails, while the extracted features provide supports for usability. Extensive experiments have been conducted to validates our intuition. The model trained by DAMIA has a negligible footprint to the usability. Our experiment also excludes factors that may hinder the performance of DAMIA, providing a potential guideline to vendors and researchers to benefit from our solution in a timely manner.
Neural networks are susceptible to data inference attacks such as the model inversion attack and the membership inference attack, where the attacker could infer the reconstruction and the membership of a data sample from the confidence scores predicted by the target classifier. In this paper, we propose a unified approach, namely purification framework, to defend data inference attacks. It purifies the confidence score vectors predicted by the target classifier by reducing their dispersion. The purifier can be further specialized in defending a particular attack via adversarial learning. We evaluate our approach on benchmark datasets and classifiers. We show that when the purifier is dedicated to one attack, it naturally defends the other one, which empirically demonstrates the connection between the two attacks. The purifier can effectively defend both attacks. For example, it can reduce the membership inference accuracy by up to 15% and increase the model inversion error by a factor of up to 4. Besides, it incurs less than 0.4% classification accuracy drop and less than 5.5% distortion to the confidence scores.
The right to be forgotten states that a data owner has the right to erase their data from an entity storing it. In the context of machine learning (ML), the right to be forgotten requires an ML model owner to remove the data owner's data from the training set used to build the ML model, a process known as machine unlearning. While originally designed to protect the privacy of the data owner, we argue that machine unlearning may leave some imprint of the data in the ML model and thus create unintended privacy risks. In this paper, we perform the first study on investigating the unintended information leakage caused by machine unlearning. We propose a novel membership inference attack that leverages the different outputs of an ML model's two versions to infer whether a target sample is part of the training set of the original model but out of the training set of the corresponding unlearned model. Our experiments demonstrate that the proposed membership inference attack achieves strong performance. More importantly, we show that our attack in multiple cases outperforms the classical membership inference attack on the original ML model, which indicates that machine unlearning can have counterproductive effects on privacy. We notice that the privacy degradation is especially significant for well-generalized ML models where classical membership inference does not perform well. We further investigate four mechanisms to mitigate the newly discovered privacy risks and show that releasing the predicted label only, temperature scaling, and differential privacy are effective. We believe that our results can help improve privacy protection in practical implementations of machine unlearning. Our code is available at https://github.com/MinChen00/UnlearningLeaks.
Federated learning (FL) is a decentralized model training framework that aims to merge isolated data islands while maintaining data privacy. However, recent studies have revealed that Generative Adversarial Network (GAN) based attacks can be employed in FL to learn the distribution of private datasets and reconstruct recognizable images. In this paper, we exploit defenses against GAN-based attacks in FL and propose a framework, Anti-GAN, to prevent attackers from learning the real distribution of the victim's data. The core idea of Anti-GAN is to manipulate the visual features of private training images to make them indistinguishable to human eyes even restored by attackers. Specifically, Anti-GAN projects the private dataset onto a GAN's generator and combines the generated fake images with the actual images to create the training dataset, which is then used for federated model training. The experimental results demonstrate that Anti-GAN is effective in preventing attackers from learning the distribution of private images while causing minimal harm to the accuracy of the federated model.
Machine learning models are prone to memorizing sensitive data, making them vulnerable to membership inference attacks in which an adversary aims to guess if an input sample was used to train the model. In this paper, we show that prior work on membership inference attacks may severely underestimate the privacy risks by relying solely on training custom neural network classifiers to perform attacks and focusing only on the aggregate results over data samples, such as the attack accuracy. To overcome these limitations, we first propose to benchmark membership inference privacy risks by improving existing non-neural network based inference attacks and proposing a new inference attack method based on a modification of prediction entropy. We also propose benchmarks for defense mechanisms by accounting for adaptive adversaries with knowledge of the defense and also accounting for the trade-off between model accuracy and privacy risks. Using our benchmark attacks, we demonstrate that existing defense approaches are not as effective as previously reported. Next, we introduce a new approach for fine-grained privacy analysis by formulating and deriving a new metric called the privacy risk score. Our privacy risk score metric measures an individual sample's likelihood of being a training member, which allows an adversary to identify samples with high privacy risks and perform attacks with high confidence. We experimentally validate the effectiveness of the privacy risk score and demonstrate that the distribution of privacy risk score across individual samples is heterogeneous. Finally, we perform an in-depth investigation for understanding why certain samples have high privacy risks, including correlations with model sensitivity, generalization error, and feature embeddings. Our work emphasizes the importance of a systematic and rigorous evaluation of privacy risks of machine learning models.
Machine learning (ML) models have been shown to be vulnerable to Membership Inference Attacks (MIA), which infer the membership of a given data point in the target dataset by observing the prediction output of the ML model. While the key factors for the success of MIA have not yet been fully understood, existing defense mechanisms such as using L2 regularization \cite{10shokri2017membership} and dropout layers \cite{salem2018ml} take only the model's overfitting property into consideration. In this paper, we provide an empirical analysis of the impact of both the data and ML model properties on the vulnerability of ML techniques to MIA. Our results reveal the relationship between MIA accuracy and properties of the dataset and training model in use. In particular, we show that the size of shadow dataset, the class and feature balance and the entropy of the target dataset, the configurations and fairness of the training model are the most influential factors. Based on those experimental findings, we conclude that along with model overfitting, multiple properties jointly contribute to MIA success instead of any single property. Building on our experimental findings, we propose using those data and model properties as regularizers to protect ML models against MIA. Our results show that the proposed defense mechanisms can reduce the MIA accuracy by up to 25\% without sacrificing the ML model prediction utility.
Iterative clustering algorithms help us to learn the insights behind the data. Unfortunately, this may allow adversaries to infer the privacy of individuals with some background knowledge. In the worst case, the adversaries know the centroids of an arbitrary iteration and the information of n-1 out of n items. To protect individual privacy against such an inference attack, preserving differential privacy (DP) for the iterative clustering algorithms has been extensively studied in the interactive settings. However, existing interactive differentially private clustering algorithms suffer from a non-convergence problem, i.e., these algorithms may not terminate without a predefined number of iterations. This problem severely impacts the clustering quality and the efficiency of a differentially private algorithm. To resolve this problem, in this paper, we propose a novel differentially private clustering framework in the interactive settings which controls the orientation of the movement of the centroids over the iterations to ensure the convergence by injecting DP noise in a selected area. We prove that, in the expected case, algorithm under our framework converges in at most twice the iterations of Lloyd's algorithm. We perform experimental evaluations on real-world datasets to show that our algorithm outperforms the state-of-the-art of the interactive differentially private clustering algorithms with guaranteed convergence and better clustering quality to meet the same DP requirement.
Machine learning models can leak information regarding the dataset they have trained. In this paper, we present the first membership inference attack against black-boxed object detection models that determines whether the given data records are used in the training. To attack the object detection model, we devise a novel method named as called a canvas method, in which predicted bounding boxes are drawn on an empty image for the attack model input. Based on the experiments, we successfully reveal the membership status of privately sensitive data trained using one-stage and two-stage detection models. We then propose defense strategies and also conduct a transfer attack between the models and datasets. Our results show that object detection models are also vulnerable to inference attacks like other models.
Attacks that aim to identify the training data of public neural networks represent a severe threat to the privacy of individuals participating in the training data set. A possible protection is offered by anonymization of the training data or training function with differential privacy. However, data scientists can choose between local and central differential privacy and need to select meaningful privacy parameters $\epsilon$ which is challenging for non-privacy experts. We empirically compare local and central differential privacy mechanisms under white- and black-box membership inference to evaluate their relative privacy-accuracy trade-offs. We experiment with several datasets and show that this trade-off is similar for both types of mechanisms. This suggests that local differential privacy is a sound alternative to central differential privacy for differentially private deep learning, since small $\epsilon$ in central differential privacy and large $\epsilon$ in local differential privacy result in similar membership inference attack risk.
Introducing noise in the training of machine learning systems is a powerful way to protect individual privacy via differential privacy guarantees, but comes at a cost to utility. This work looks at whether the inherent randomness of stochastic gradient descent (SGD) could contribute to privacy, effectively reducing the amount of \emph{additional} noise required to achieve a given privacy guarantee. We conduct a large-scale empirical study to examine this question. Training a grid of over 120,000 models across four datasets (tabular and images) on convex and non-convex objectives, we demonstrate that the random seed has a larger impact on model weights than any individual training example. We test the distribution over weights induced by the seed, finding that the simple convex case can be modelled with a multivariate Gaussian posterior, while neural networks exhibit multi-modal and non-Gaussian weight distributions. By casting convex SGD as a Gaussian mechanism, we then estimate an `intrinsic' data-dependent $\epsilon_i(\mathcal{D})$, finding values as low as 6.3, dropping to 1.9 using empirical estimates. We use a membership inference attack to estimate $\epsilon$ for non-convex SGD and demonstrate that hiding the random seed from the adversary results in a statistically significant reduction in attack performance, corresponding to a reduction in the effective $\epsilon$. These results provide empirical evidence that SGD exhibits appreciable variability relative to its dataset sensitivity, and this `intrinsic noise' has the potential to be leveraged to improve the utility of privacy-preserving machine learning.
Membership inference attacks seek to infer the membership of individual training instances of a privately trained model. This paper presents a membership privacy analysis and evaluation system, called MPLens, with three unique contributions. First, through MPLens, we demonstrate how membership inference attack methods can be leveraged in adversarial machine learning. Second, through MPLens, we highlight how the vulnerability of pre-trained models under membership inference attack is not uniform across all classes, particularly when the training data itself is skewed. We show that risk from membership inference attacks is routinely increased when models use skewed training data. Finally, we investigate the effectiveness of differential privacy as a mitigation technique against membership inference attacks. We discuss the trade-offs of implementing such a mitigation strategy with respect to the model complexity, the learning task complexity, the dataset complexity and the privacy parameter settings. Our empirical results reveal that (1) minority groups within skewed datasets display increased risk for membership inference and (2) differential privacy presents many challenging trade-offs as a mitigation technique to membership inference risk.
It is critical to understand the privacy and robustness vulnerabilities of machine learning models, as their implementation expands in scope. In membership inference attacks, adversaries can determine whether a particular set of data was used in training, putting the privacy of the data at risk. Existing work has mostly focused on image related tasks; we generalize this type of attack to speaker identification on audio samples. We demonstrate attack precision of 85.9\% and recall of 90.8\% for LibriSpeech, and 78.3\% precision and 90.7\% recall for VOiCES (Voices Obscured in Complex Environmental Settings). We find that implementing defenses such as prediction obfuscation, defensive distillation or adversarial training, can reduce attack accuracy to chance.
Machine Learning models, extensively used for various multimedia applications, are offered to users as a blackbox service on the Cloud on a pay-per-query basis. Such blackbox models are commercially valuable to adversaries, making them vulnerable to extraction attacks to reverse engineer the proprietary model thereby violating the model privacy and Intellectual Property. Here, the adversary first extracts the model architecture or hyperparameters through side channel leakage, followed by stealing the functionality of the target model by training the reconstructed architecture on a synthetic dataset. While the attacks proposed in literature are empirical, there is a need for a theoretical framework to measure the information leaked under such extraction attacks. To this extent, in this work, we propose a novel probabilistic framework, Airavata, to estimate the information leakage in such model extraction attacks. This framework captures the fact that extracting the exact target model is difficult due to experimental uncertainty while inferring model hyperparameters and stochastic nature of training to steal the target model functionality. Specifically, we use Bayesian Networks to capture uncertainty in estimating the target model under various extraction attacks based on the subjective notion of probability. We validate the proposed framework under different adversary assumptions commonly adopted in literature to reason about the attack efficacy. This provides a practical tool to infer actionable details about extracting blackbox models and help identify the best attack combination which maximises the knowledge extracted (or information leaked) from the target model.
Watermarking of deep neural networks (DNN) can enable their tracing once released by a data owner. In this paper, we generalize white-box watermarking algorithms for DNNs, where the data owner needs white-box access to the model to extract the watermark. White-box watermarking algorithms have the advantage that they do not impact the accuracy of the watermarked model. We propose Robust whIte-box GAn watermarking (RIGA), a novel white-box watermarking algorithm that uses adversarial training. Our extensive experiments demonstrate that the proposed watermarking algorithm not only does not impact accuracy, but also significantly improves the covertness and robustness over the current state-of-art.