Pre-training Differentially Private Models with Limited Public Data

The superior performance of large foundation models relies on the use of massive amounts of high-quality data, which often contain sensitive, private and copyrighted material that requires formal protection. While differential privacy (DP) is a prominent method to gauge the degree of security provided to the models, its application is commonly limited to the model fine-tuning stage, due to the performance degradation when applying DP during the pre-training stage. Consequently, DP is yet not capable of protecting a substantial portion of the data used during the initial pre-training process. In this work, we first provide a theoretical understanding of the efficacy of DP training by analyzing the per-iteration loss improvement. We make a key observation that DP optimizers' performance degradation can be significantly mitigated by the use of limited public data, which leads to a novel DP continual pre-training strategy. Empirically, using only 10\% of public data, our strategy can achieve DP accuracy of 41.5\% on ImageNet-21k (with $\epsilon=8$), as well as non-DP accuracy of 55.7\% and and 60.0\% on downstream tasks Places365 and iNaturalist-2021, respectively, on par with state-of-the-art standard pre-training and substantially outperforming existing DP pre-trained models.

Boundary Exploration for Bayesian Optimization With Unknown Physical Constraints

Bayesian optimization has been successfully applied to optimize black-box functions where the number of evaluations is severely limited. However, in many real-world applications, it is hard or impossible to know in advance which designs are feasible due to some physical or system limitations. These issues lead to an even more challenging problem of optimizing an unknown function with unknown constraints. In this paper, we observe that in such scenarios optimal solution typically lies on the boundary between feasible and infeasible regions of the design space, making it considerably more difficult than that with interior optima. Inspired by this observation, we propose BE-CBO, a new Bayesian optimization method that efficiently explores the boundary between feasible and infeasible designs. To identify the boundary, we learn the constraints with an ensemble of neural networks that outperform the standard Gaussian Processes for capturing complex boundaries. Our method demonstrates superior performance against state-of-the-art methods through comprehensive experiments on synthetic and real-world benchmarks.

Pre-trained Trojan Attacks for Visual Recognition

Pre-trained vision models (PVMs) have become a dominant component due to their exceptional performance when fine-tuned for downstream tasks. However, the presence of backdoors within PVMs poses significant threats. Unfortunately, existing studies primarily focus on backdooring PVMs for the classification task, neglecting potential inherited backdoors in downstream tasks such as detection and segmentation. In this paper, we propose the Pre-trained Trojan attack, which embeds backdoors into a PVM, enabling attacks across various downstream vision tasks. We highlight the challenges posed by cross-task activation and shortcut connections in successful backdoor attacks. To achieve effective trigger activation in diverse tasks, we stylize the backdoor trigger patterns with class-specific textures, enhancing the recognition of task-irrelevant low-level features associated with the target class in the trigger pattern. Moreover, we address the issue of shortcut connections by introducing a context-free learning pipeline for poison training. In this approach, triggers without contextual backgrounds are directly utilized as training data, diverging from the conventional use of clean images. Consequently, we establish a direct shortcut from the trigger to the target class, mitigating the shortcut connection issue. We conducted extensive experiments to thoroughly validate the effectiveness of our attacks on downstream detection and segmentation tasks. Additionally, we showcase the potential of our approach in more practical scenarios, including large vision models and 3D object detection in autonomous driving. This paper aims to raise awareness of the potential threats associated with applying PVMs in practical scenarios. Our codes will be available upon paper publication.

Differentially Private SGD Without Clipping Bias: An Error-Feedback Approach

Differentially Private Stochastic Gradient Descent with gradient clipping (DPSGD-GC) is a powerful tool for training deep learning models using sensitive data, providing both a solid theoretical privacy guarantee and high efficiency. However, using DPSGD-GC to ensure Differential Privacy (DP) comes at the cost of model performance degradation due to DP noise injection and gradient clipping. Existing research has extensively analyzed the theoretical convergence of DPSGD-GC, and has shown that it only converges when using large clipping thresholds that are dependent on problem-specific parameters. Unfortunately, these parameters are often unknown in practice, making it hard to choose the optimal clipping threshold. Therefore, in practice, DPSGD-GC suffers from degraded performance due to the {\it constant} bias introduced by the clipping. In our work, we propose a new error-feedback (EF) DP algorithm as an alternative to DPSGD-GC, which not only offers a diminishing utility bound without inducing a constant clipping bias, but more importantly, it allows for an arbitrary choice of clipping threshold that is independent of the problem. We establish an algorithm-specific DP analysis for our proposed algorithm, providing privacy guarantees based on R{\'e}nyi DP. Additionally, we demonstrate that under mild conditions, our algorithm can achieve nearly the same utility bound as DPSGD without gradient clipping. Our empirical results on Cifar-10/100 and E2E datasets, show that the proposed algorithm achieves higher accuracies than DPSGD while maintaining the same level of DP guarantee.

On semi-supervised estimation using exponential tilt mixture models

Consider a semi-supervised setting with a labeled dataset of binary responses and predictors and an unlabeled dataset with only the predictors. Logistic regression is equivalent to an exponential tilt model in the labeled population. For semi-supervised estimation, we develop further analysis and understanding of a statistical approach using exponential tilt mixture (ETM) models and maximum nonparametric likelihood estimation, while allowing that the class proportions may differ between the unlabeled and labeled data. We derive asymptotic properties of ETM-based estimation and demonstrate improved efficiency over supervised logistic regression in a random sampling setup and an outcome-stratified sampling setup previously used. Moreover, we reconcile such efficiency improvement with the existing semiparametric efficiency theory when the class proportions in the unlabeled and labeled data are restricted to be the same. We also provide a simulation study to numerically illustrate our theoretical findings.

A new method of modeling the multi-stage decision-making process of CRT using machine learning with uncertainty quantification

Aims. The purpose of this study is to create a multi-stage machine learning model to predict cardiac resynchronization therapy (CRT) response for heart failure (HF) patients. This model exploits uncertainty quantification to recommend additional collection of single-photon emission computed tomography myocardial perfusion imaging (SPECT MPI) variables if baseline clinical variables and features from electrocardiogram (ECG) are not sufficient. Methods. 218 patients who underwent rest-gated SPECT MPI were enrolled in this study. CRT response was defined as an increase in left ventricular ejection fraction (LVEF) > 5% at a 6 month follow-up. A multi-stage ML model was created by combining two ensemble models. Results. The response rate for CRT was 55.5% (n = 121) with overall male gender 61.0% (n = 133), an average age of 62.0, and LVEF of 27.7. The multi-stage model performed similarly to Ensemble 2 (which utilized the additional SPECT data) with AUC of 0.75 vs. 0.77, accuracy of 0.71 vs. 0.69, sensitivity of 0.70 vs. 0.72, and specificity 0.72 vs. 0.65, respectively. However, the multi-stage model only required SPECT MPI data for 52.7% of the patients across all folds. Conclusions. By using rule-based logic stemming from uncertainty quantification, the multi-stage model was able to reduce the need for additional SPECT MPI data acquisition without sacrificing performance.

A new method using deep transfer learning on ECG to predict the response to cardiac resynchronization therapy

Background: Cardiac resynchronization therapy (CRT) has emerged as an effective treatment for heart failure patients with electrical dyssynchrony. However, accurately predicting which patients will respond to CRT remains a challenge. This study explores the application of deep transfer learning techniques to train a predictive model for CRT response. Methods: In this study, the short-time Fourier transform (STFT) technique was employed to transform ECG signals into two-dimensional images. A transfer learning approach was then applied on the MIT-BIT ECG database to pre-train a convolutional neural network (CNN) model. The model was fine-tuned to extract relevant features from the ECG images, and then tested on our dataset of CRT patients to predict their response. Results: Seventy-one CRT patients were enrolled in this study. The transfer learning model achieved an accuracy of 72% in distinguishing responders from non-responders in the local dataset. Furthermore, the model showed good sensitivity (0.78) and specificity (0.79) in identifying CRT responders. The performance of our model outperformed clinic guidelines and traditional machine learning approaches. Conclusion: The utilization of ECG images as input and leveraging the power of transfer learning allows for improved accuracy in identifying CRT responders. This approach offers potential for enhancing patient selection and improving outcomes of CRT.

A new method using deep learning to predict the response to cardiac resynchronization therapy

Background. Clinical parameters measured from gated single-photon emission computed tomography myocardial perfusion imaging (SPECT MPI) have value in predicting cardiac resynchronization therapy (CRT) patient outcomes, but still show limitations. The purpose of this study is to combine clinical variables, features from electrocardiogram (ECG), and parameters from assessment of cardiac function with polarmaps from gated SPECT MPI through deep learning (DL) to predict CRT response. Methods. 218 patients who underwent rest gated SPECT MPI were enrolled in this study. CRT response was defined as an increase in left ventricular ejection fraction (LVEF) > 5% at a 6-month follow up. A DL model was constructed by combining a pre-trained VGG16 module and a multilayer perceptron. Two modalities of data were input to the model: polarmap images from SPECT MPI and tabular data from clinical features and ECG parameters. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to the VGG16 module to provide explainability for the polarmaps. For comparison, four machine learning (ML) models were trained using only the tabular features. Results. Modeling was performed on 218 patients who underwent CRT implantation with a response rate of 55.5% (n = 121). The DL model demonstrated average AUC (0.83), accuracy (0.73), sensitivity (0.76), and specificity (0.69) surpassing the ML models and guideline criteria. Guideline recommendations presented accuracy (0.53), sensitivity (0.75), and specificity (0.26). Conclusions. The DL model outperformed the ML models, showcasing the additional predictive benefit of utilizing SPECT MPI polarmaps. Incorporating additional patient data directly in the form of medical imagery can improve CRT response prediction.

Persistently Trained, Diffusion-assisted Energy-based Models

Maximum likelihood (ML) learning for energy-based models (EBMs) is challenging, partly due to non-convergence of Markov chain Monte Carlo.Several variations of ML learning have been proposed, but existing methods all fail to achieve both post-training image generation and proper density estimation. We propose to introduce diffusion data and learn a joint EBM, called diffusion assisted-EBMs, through persistent training (i.e., using persistent contrastive divergence) with an enhanced sampling algorithm to properly sample from complex, multimodal distributions. We present results from a 2D illustrative experiment and image experiments and demonstrate that, for the first time for image data, persistently trained EBMs can {\it simultaneously} achieve long-run stability, post-training image generation, and superior out-of-distribution detection.

GLASU: A Communication-Efficient Algorithm for Federated Learning with Vertically Distributed Graph Data

Vertical federated learning (VFL) is a distributed learning paradigm, where computing clients collectively train a model based on the partial features of the same set of samples they possess. Current research on VFL focuses on the case when samples are independent, but it rarely addresses an emerging scenario when samples are interrelated through a graph. For graph-structured data, graph neural networks (GNNs) are competitive machine learning models, but a naive implementation in the VFL setting causes a significant communication overhead. Moreover, the analysis of the training is faced with a challenge caused by the biased stochastic gradients. In this paper, we propose a model splitting method that splits a backbone GNN across the clients and the server and a communication-efficient algorithm, GLASU, to train such a model. GLASU adopts lazy aggregation and stale updates to skip aggregation when evaluating the model and skip feature exchanges during training, greatly reducing communication. We offer a theoretical analysis and conduct extensive numerical experiments on real-world datasets, showing that the proposed algorithm effectively trains a GNN model, whose performance matches that of the backbone GNN when trained in a centralized manner.