Search

Large Language Models (LLMs) enable a new ecosystem with many downstream applications, called LLM applications, with different natural language processing tasks. The functionality and performance of an LLM application highly depend on its system prompt, which instructs the backend LLM on what task to perform. Therefore, an LLM application developer often keeps a system prompt confidential to protect its intellectual property. As a result, a natural attack, called prompt leaking, is to steal the system prompt from an LLM application, which compromises the developer's intellectual property. Existing prompt leaking attacks primarily rely on manually crafted queries, and thus achieve limited effectiveness. In this paper, we design a novel, closed-box prompt leaking attack framework, called PLeak, to optimize an adversarial query such that when the attacker sends it to a target LLM application, its response reveals its own system prompt. We formulate finding such an adversarial query as an optimization problem and solve it with a gradient-based method approximately. Our key idea is to break down the optimization goal by optimizing adversary queries for system prompts incrementally, i.e., starting from the first few tokens of each system prompt step by step until the entire length of the system prompt. We evaluate PLeak in both offline settings and for real-world LLM applications, e.g., those on Poe, a popular platform hosting such applications. Our results show that PLeak can effectively leak system prompts and significantly outperforms not only baselines that manually curate queries but also baselines with optimized queries that are modified and adapted from existing jailbreaking attacks. We responsibly reported the issues to Poe and are still waiting for their response. Our implementation is available at this repository: https://github.com/BHui97/PLeak., Comment: To appear in the Proceedings of The ACM Conference on Computer and Communications Security (CCS), 2024

This paper investigates to what degree and magnitude tradeoffs exist between utility, fairness and attribute privacy in computer vision. Regarding privacy, we look at this important problem specifically in the context of attribute inference attacks, a less addressed form of privacy. To create a variety of models with different preferences, we use adversarial methods to intervene on attributes relating to fairness and privacy. We see that that certain tradeoffs exist between fairness and utility, privacy and utility, and between privacy and fairness. The results also show that those tradeoffs and interactions are more complex and nonlinear between the three goals than intuition would suggest.

Federated learning (FL) allows multiple clients to collaboratively train a deep learning model. One major challenge of FL is when data distribution is heterogeneous, i.e., differs from one client to another. Existing personalized FL algorithms are only applicable to narrow cases, e.g., one or two data classes per client, and therefore they do not satisfactorily address FL under varying levels of data heterogeneity. In this paper, we propose a novel framework, called DisTrans, to improve FL performance (i.e., model accuracy) via train and test-time distributional transformations along with a double-input-channel model structure. DisTrans works by optimizing distributional offsets and models for each FL client to shift their data distribution, and aggregates these offsets at the FL server to further improve performance in case of distributional heterogeneity. Our evaluation on multiple benchmark datasets shows that DisTrans outperforms state-of-the-art FL methods and data augmentation methods under various settings and different degrees of client distributional heterogeneity., Comment: In the Proceedings of European Conference on Computer Vision (ECCV), 2022

Designing machine learning algorithms that are accurate yet fair, not discriminating based on any sensitive attribute, is of paramount importance for society to accept AI for critical applications. In this article, we propose a novel fair representation learning method termed the R\'enyi Fair Information Bottleneck Method (RFIB) which incorporates constraints for utility, fairness, and compactness (compression) of representation, and apply it to image and tabular data classification. A key attribute of our approach is that we consider - in contrast to most prior work - both demographic parity and equalized odds as fairness constraints, allowing for a more nuanced satisfaction of both criteria. Leveraging a variational approach, we show that our objectives yield a loss function involving classical Information Bottleneck (IB) measures and establish an upper bound in terms of two R\'enyi measures of order $\alpha$ on the mutual information IB term measuring compactness between the input and its encoded embedding. We study the influence of the $\alpha$ parameter as well as two other tunable IB parameters on achieving utility/fairness trade-off goals, and show that the $\alpha$ parameter gives an additional degree of freedom that can be used to control the compactness of the representation. Experimenting on three different image datasets (EyePACS, CelebA, and FairFace) and two tabular datasets (Adult and COMPAS), using both binary and categorical sensitive attributes, we show that on various utility, fairness, and compound utility/fairness metrics RFIB outperforms current state-of-the-art approaches., Comment: To appear in the IEEE Transactions on Information Forensics & Security

We develop a novel method for ensuring fairness in machine learning which we term as the Renyi Fair Information Bottleneck (RFIB). We consider two different fairness constraints - demographic parity and equalized odds - for learning fair representations and derive a loss function via a variational approach that uses Renyi's divergence with its tunable parameter $\alpha$ and that takes into account the triple constraints of utility, fairness, and compactness of representation. We then evaluate the performance of our method for image classification using the EyePACS medical imaging dataset, showing it outperforms competing state of the art techniques with performance measured using a variety of compound utility/fairness metrics, including accuracy gap and Rawls' minimal accuracy., Comment: To appear in the Proceedings of CWIT'22

We tackle here a specific, still not widely addressed aspect, of AI robustness, which consists of seeking invariance / insensitivity of model performance to hidden factors of variations in the data. Towards this end, we employ a two step strategy that a) does unsupervised discovery, via generative models, of sensitive factors that cause models to under-perform, and b) intervenes models to make their performance invariant to these sensitive factors' influence. We consider 3 separate interventions for robustness, including: data augmentation, semantic consistency, and adversarial alignment. We evaluate our method using metrics that measure trade offs between invariance (insensitivity) and overall performance (utility) and show the benefits of our method for 3 settings (unsupervised, semi-supervised and generalization).

Skin lesions can be an early indicator of a wide range of infectious and other diseases. The use of deep learning (DL) models to diagnose skin lesions has great potential in assisting clinicians with prescreening patients. However, these models often learn biases inherent in training data, which can lead to a performance gap in the diagnosis of people with light and/or dark skin tones. To the best of our knowledge, limited work has been done on identifying, let alone reducing, model bias in skin disease classification and segmentation. In this paper, we examine DL fairness and demonstrate the existence of bias in classification and segmentation models for subpopulations with darker skin tones compared to individuals with lighter skin tones, for specific diseases including Lyme, Tinea Corporis and Herpes Zoster. Then, we propose a novel preprocessing, data alteration method, called EdgeMixup, to improve model fairness with a linear combination of an input skin lesion image and a corresponding a predicted edge detection mask combined with color saturation alteration. For the task of skin disease classification, EdgeMixup outperforms much more complex competing methods such as adversarial approaches, achieving a 10.99% reduction in accuracy gap between light and dark skin tone samples, and resulting in 8.4% improved performance for an underrepresented subpopulation.

Perturbation-based attacks, while not physically realizable, have been the main emphasis of adversarial machine learning (ML) research. Patch-based attacks by contrast are physically realizable, yet most work has focused on 2D domain with recent forays into 3D. Characterizing the robustness properties of patch attacks and their invariance to 3D pose is important, yet not fully elucidated, and is the focus of this paper. To this end, several contributions are made here: A) we develop a new metric called mean Attack Success over Transformations (mAST) to evaluate patch attack robustness and invariance; and B), we systematically assess robustness of patch attacks to 3D position and orientation for various conditions; in particular, we conduct a sensitivity analysis which provides important qualitative insights into attack effectiveness as a function of the 3D pose of a patch relative to the camera (rotation, translation) and sets forth some properties for patch attack 3D invariance; and C), we draw novel qualitative conclusions including: 1) we demonstrate that for some 3D transformations, namely rotation and loom, increasing the training distribution support yields an increase in patch success over the full range at test time. 2) We provide new insights into the existence of a fundamental cutoff limit in patch attack effectiveness that depends on the extent of out-of-plane rotation angles. These findings should collectively guide future design of 3D patch attacks and defenses.

Assured AI in unrestricted settings is a critical problem. Our framework addresses AI assurance challenges lying at the intersection of domain adaptation, fairness, and counterfactuals analysis, operating via the discovery and intervention on factors of variations in data (e.g. weather or illumination conditions) that significantly affect the robustness of AI models. Robustness is understood here as insensitivity of the model performance to variations in sensitive factors. Sensitive factors are traditionally set in a supervised setting, whereby factors are known a-priori (e.g. for fairness this could be factors like sex or race). In contrast, our motivation is real-life scenarios where less, or nothing, is actually known a-priori about certain factors that cause models to fail. This leads us to consider various settings (unsupervised, domain generalization, semi-supervised) that correspond to different degrees of incomplete knowledge about those factors. Therefore, our two step approach works by a) discovering sensitive factors that cause AI systems to fail in a unsupervised fashion, and then b) intervening models to lessen these factor's influence. Our method considers 3 interventions consisting of Augmentation, Coherence, and Adversarial Interventions (ACAI). We demonstrate the ability for interventions on discovered/source factors to generalize to target/real factors. We also demonstrate how adaptation to real factors of variations can be performed in the semi-supervised case where some target factor labels are known, via automated intervention selection. Experiments show that our approach improves on baseline models, with regard to achieving optimal utility vs. sensitivity/robustness tradeoffs.

Membership inference (MI) attacks affect user privacy by inferring whether given data samples have been used to train a target learning model, e.g., a deep neural network. There are two types of MI attacks in the literature, i.e., these with and without shadow models. The success of the former heavily depends on the quality of the shadow model, i.e., the transferability between the shadow and the target; the latter, given only blackbox probing access to the target model, cannot make an effective inference of unknowns, compared with MI attacks using shadow models, due to the insufficient number of qualified samples labeled with ground truth membership information. In this paper, we propose an MI attack, called BlindMI, which probes the target model and extracts membership semantics via a novel approach, called differential comparison. The high-level idea is that BlindMI first generates a dataset with nonmembers via transforming existing samples into new samples, and then differentially moves samples from a target dataset to the generated, non-member set in an iterative manner. If the differential move of a sample increases the set distance, BlindMI considers the sample as non-member and vice versa. BlindMI was evaluated by comparing it with state-of-the-art MI attack algorithms. Our evaluation shows that BlindMI improves F1-score by nearly 20% when compared to state-of-the-art on some datasets, such as Purchase-50 and Birds-200, in the blind setting where the adversary does not know the target model's architecture and the target dataset's ground truth labels. We also show that BlindMI can defeat state-of-the-art defenses.

Searching for small objects in large images is a task that is both challenging for current deep learning systems and important in numerous real-world applications, such as remote sensing and medical imaging. Thorough scanning of very large images is computationally expensive, particularly at resolutions sufficient to capture small objects. The smaller an object of interest, the more likely it is to be obscured by clutter or otherwise deemed insignificant. We examine these issues in the context of two complementary problems: closed-set object detection and open-set target search. First, we present a method for predicting pixel-level objectness from a low resolution gist image, which we then use to select regions for performing object detection locally at high resolution. This approach has the benefit of not being fixed to a predetermined grid, thereby requiring fewer costly high-resolution glimpses than existing methods. Second, we propose a novel strategy for open-set visual search that seeks to find all instances of a target class which may be previously unseen and is defined by a single image. We interpret both detection problems through a probabilistic, Bayesian lens, whereby the objectness maps produced by our method serve as priors in a maximum-a-posteriori approach to the detection step. We evaluate the end-to-end performance of both the combination of our patch selection strategy with this target search approach and the combination of our patch selection strategy with standard object detection methods. Both elements of our approach are seen to significantly outperform baseline strategies.

Whilst adversarial attack detection has received considerable attention, it remains a fundamentally challenging problem from two perspectives. First, while threat models can be well-defined, attacker strategies may still vary widely within those constraints. Therefore, detection should be considered as an open-set problem, standing in contrast to most current detection approaches. These methods take a closed-set view and train binary detectors, thus biasing detection toward attacks seen during detector training. Second, limited information is available at test time and typically confounded by nuisance factors including the label and underlying content of the image. We address these challenges via a novel strategy based on random subspace analysis. We present a technique that utilizes properties of random projections to characterize the behavior of clean and adversarial examples across a diverse set of subspaces. The self-consistency (or inconsistency) of model activations is leveraged to discern clean from adversarial examples. Performance evaluations demonstrate that our technique ($AUC\in[0.92, 0.98]$) outperforms competing detection strategies ($AUC\in[0.30,0.79]$), while remaining truly agnostic to the attack strategy (for both targeted/untargeted attacks). It also requires significantly less calibration data (composed only of clean examples) than competing approaches to achieve this performance.

We examine progress in the use of AI for detecting skin lesions, with particular emphasis on the erythema migrans rash of acute Lyme disease, and other lesions, such as those from conditions like herpes zoster (shingles), tinea corporis, erythema multiforme, cellulitis, insect bites, or tick bites. We discuss important challenges for these applications, in particular the problems of AI bias regarding the lack of skin images in dark skinned individuals, being able to accurately detect, delineate, and segment lesions or regions of interest compared to normal skin in images, and low shot learning (addressing classification with a paucity of training images). Solving these problems ranges from being highly desirable requirements -- e.g. for delineation, which may be useful to disambiguate between similar types of lesions, and perform improved diagnostics -- or required, as is the case for AI de-biasing, to allow for the deployment of fair AI techniques in the clinic for skin lesion analysis. For the problem of low shot learning in particular, we report skin analysis algorithms that gracefully degrade and still perform well at low shots, when compared to baseline algorithms: when using a little as 10 training exemplars per class, the baseline DL algorithm performance significantly degrades, with accuracy of 56.41%, close to chance, whereas the best performing low shot algorithm yields an accuracy of 85.26%.

We investigate the use of parametrized families of information-theoretic measures to generalize the loss functions of generative adversarial networks (GANs) with the objective of improving performance. A new generator loss function, called least $k$th-order GAN (L$k$GAN), is first introduced, generalizing the least squares GANs (LSGANs) by using a $k$th order absolute error distortion measure with $k \geq 1$ (which recovers the LSGAN loss function when $k=2$). It is shown that minimizing this generalized loss function under an (unconstrained) optimal discriminator is equivalent to minimizing the $k$th-order Pearson-Vajda divergence. Another novel GAN generator loss function is next proposed in terms of R\'{e}nyi cross-entropy functionals with order $\alpha >0$, $\alpha\neq 1$. It is demonstrated that this R\'{e}nyi-centric generalized loss function, which provably reduces to the original GAN loss function as $\alpha\to1$, preserves the equilibrium point satisfied by the original GAN based on the Jensen-R\'{e}nyi divergence, a natural extension of the Jensen-Shannon divergence. Experimental results indicate that the proposed loss functions, applied to the MNIST and CelebA datasets, under both DCGAN and StyleGAN architectures, confer performance benefits by virtue of the extra degrees of freedom provided by the parameters $k$ and $\alpha$, respectively. More specifically, experiments show improvements with regard to the quality of the generated images as measured by the Fr\'echet Inception Distance (FID) score and training stability. While it was applied to GANs in this study, the proposed approach is generic and can be used in other applications of information theory to deep learning, e.g., the issues of fairness or privacy in artificial intelligence.

We focus on the development of effective adversarial patch attacks and -- for the first time -- jointly address the antagonistic objectives of attack success and obtrusiveness via the design of novel semi-transparent patches. This work is motivated by our pursuit of a systematic performance analysis of patch attack robustness with regard to geometric transformations. Specifically, we first elucidate a) key factors underpinning patch attack success and b) the impact of distributional shift between training and testing/deployment when cast under the Expectation over Transformation (EoT) formalism. By focusing our analysis on three principal classes of transformations (rotation, scale, and location), our findings provide quantifiable insights into the design of effective patch attacks and demonstrate that scale, among all factors, significantly impacts patch attack success. Working from these findings, we then focus on addressing how to overcome the principal limitations of scale for the deployment of attacks in real physical settings: namely the obtrusiveness of large patches. Our strategy is to turn to the novel design of irregularly-shaped, semi-transparent partial patches which we construct via a new optimization process that jointly addresses the antagonistic goals of mitigating obtrusiveness and maximizing effectiveness. Our study -- we hope -- will help encourage more focus in the community on the issues of obtrusiveness, scale, and success in patch attacks.

This study evaluated generative methods to potentially mitigate AI bias when diagnosing diabetic retinopathy (DR) resulting from training data imbalance, or domain generalization which occurs when deep learning systems (DLS) face concepts at test/inference time they were not initially trained on. The public domain Kaggle-EyePACS dataset (88,692 fundi and 44,346 individuals, originally diverse for ethnicity) was modified by adding clinician-annotated labels and constructing an artificial scenario of data imbalance and domain generalization by disallowing training (but not testing) exemplars for images of retinas with DR warranting referral (DR-referable) and from darker-skin individuals, who presumably have greater concentration of melanin within uveal melanocytes, on average, contributing to retinal image pigmentation. A traditional/baseline diagnostic DLS was compared against new DLSs that would use training data augmented via generative models for debiasing. Accuracy (95% confidence intervals [CI]) of the baseline diagnostics DLS for fundus images of lighter-skin individuals was 73.0% (66.9%, 79.2%) vs. darker-skin of 60.5% (53.5%, 67.3%), demonstrating bias/disparity (delta=12.5%) (Welch t-test t=2.670, P=.008) in AI performance across protected subpopulations. Using novel generative methods for addressing missing subpopulation training data (DR-referable darker-skin) achieved instead accuracy, for lighter-skin, of 72.0% (65.8%, 78.2%), and for darker-skin, of 71.5% (65.2%,77.8%), demonstrating closer parity (delta=0.5%) in accuracy across subpopulations (Welch t-test t=0.111, P=.912). Findings illustrate how data imbalance and domain generalization can lead to disparity of accuracy across subpopulations, and show that novel generative methods of synthetic fundus images may play a role for debiasing AI., Comment: Accepted for publication at journal of Translational Vision Science and Technology

Our work focuses on unsupervised and generative methods that address the following goals: (a) learning unsupervised generative representations that discover latent factors controlling image semantic attributes, (b) studying how this ability to control attributes formally relates to the issue of latent factor disentanglement, clarifying related but dissimilar concepts that had been confounded in the past, and (c) developing anomaly detection methods that leverage representations learned in (a). For (a), we propose a network architecture that exploits the combination of multiscale generative models with mutual information (MI) maximization. For (b), we derive an analytical result (Lemma 1) that brings clarity to two related but distinct concepts: the ability of generative networks to control semantic attributes of images they generate, resulting from MI maximization, and the ability to disentangle latent space representations, obtained via total correlation minimization. More specifically, we demonstrate that maximizing semantic attribute control encourages disentanglement of latent factors. Using Lemma 1 and adopting MI in our loss function, we then show empirically that, for image generation tasks, the proposed approach exhibits superior performance as measured in the quality and disentanglement trade space, when compared to other state of the art methods, with quality assessed via the Frechet Inception Distance (FID), and disentanglement via mutual information gap. For (c), we design several systems for anomaly detection exploiting representations learned in (a), and demonstrate their performance benefits when compared to state-of-the-art generative and discriminative algorithms. The above contributions in representation learning have potential applications in addressing other important problems in computer vision, such as bias and privacy in AI., Comment: MIT Neural Computation 2021, Vol 33(3), pp. 802--826

Fast, collision-free motion through unknown environments remains a challenging problem for robotic systems. In these situations, the robot's ability to reason about its future motion is often severely limited by sensor field of view (FOV). By contrast, biological systems routinely make decisions by taking into consideration what might exist beyond their FOV based on prior experience. In this paper, we present an approach for predicting occupancy map representations of sensor data for future robot motions using deep neural networks. We evaluate several deep network architectures, including purely generative and adversarial models. Testing on both simulated and real environments we demonstrated performance both qualitatively and quantitatively, with SSIM similarity measure up to 0.899. We showed that it is possible to make predictions about occupied space beyond the physical robot's FOV from simulated training data. In the future, this method will allow robots to navigate through unknown environments in a faster, safer manner., Comment: 7 pages

Our objective is to evaluate the efficacy of methods that use deep learning (DL) for the automatic fine-grained segmentation of optical coherence tomography (OCT) images of the retina. OCT images from 10 patients with mild non-proliferative diabetic retinopathy were used from a public (U. of Miami) dataset. For each patient, five images were available: one image of the fovea center, two images of the perifovea, and two images of the parafovea. For each image, two expert graders each manually annotated five retinal surfaces (i.e. boundaries between pairs of retinal layers). The first grader's annotations were used as ground truth and the second grader's annotations to compute inter-operator agreement. The proposed automated approach segments images using fully convolutional networks (FCNs) together with Gaussian process (GP)-based regression as a post-processing step to improve the quality of the estimates. Using 10-fold cross validation, the performance of the algorithms is determined by computing the per-pixel unsigned error (distance) between the automated estimates and the ground truth annotations generated by the first manual grader. We compare the proposed method against five state of the art automatic segmentation techniques. The results show that the proposed methods compare favorably with state of the art techniques, resulting in the smallest mean unsigned error values and associated standard deviations, and performance is comparable with human annotation of retinal layers from OCT when there is only mild retinopathy. The results suggest that semantic segmentation using FCNs, coupled with regression-based post-processing, can effectively solve the OCT segmentation problem on par with human capabilities with mild retinopathy.

We address zero-shot (ZS) learning, building upon prior work in hierarchical classification by combining it with approaches based on semantic attribute estimation. For both non-novel and novel image classes we compare multiple formulations of the problem, starting with deep universal features in each case. We investigate the effect of using different posterior probabilities as inputs to the hierarchical classifier, comparing the performances of posteriors derived from distances to SVM classifier boundaries with those of posteriors based on semantic attribute estimation. Using a dataset consisting of 150 object classes from the ImageNet ILSVRC2012 data set, we find that the hierarchical classification method that maximizes expected reward for non-novel classes differs from the method that maximizes expected reward for novel classes. We also show that using input posteriors based on semantic attributes improves the expected reward for novel classes., Comment: 17 pages, 4 figures, extension to work published in conference proceedings of 2017 IAPR MVA Conference