skip to main content

Title: On the Need for Topology-Aware Generative Models for Manifold-Based Defenses
ML algorithms or models, especially deep neural networks (DNNs), have shown significant promise in several areas. However, recently researchers have demonstrated that ML algorithms, especially DNNs, are vulnerable to adversarial examples (slightly perturbed samples that cause mis-classification). Existence of adversarial examples has hindered deployment of ML algorithms in safety-critical sectors, such as security. Several defenses for adversarial examples exist in the literature. One of the important classes of defenses are manifold-based defenses, where a sample is “pulled back” into the data manifold before classifying. These defenses rely on the manifold assumption (data lie in a manifold of lower dimension than the input space). These defenses use a generative model to approximate the input distribution. This paper asks the following question: do the generative models used in manifold-based defenses need to be topology-aware? Our paper suggests the answer is yes. We provide theoretical and empirical evidence to support our claim.
Authors:
; ;
Award ID(s):
1740079 1750009
Publication Date:
NSF-PAR ID:
10181037
Journal Name:
8th International Conference on Learning Representations (ICLR) 2020
Sponsoring Org:
National Science Foundation
More Like this
  1. To make daily decisions, human agents devise their own "strategies" governing their mobility dynamics (e.g., taxi drivers have preferred working regions and times, and urban commuters have preferred routes and transit modes). Recent research such as generative adversarial imitation learning (GAIL) demonstrates successes in learning human decision-making strategies from their behavior data using deep neural networks (DNNs), which can accurately mimic how humans behave in various scenarios, e.g., playing video games, etc. However, such DNN-based models are "black box" models in nature, making it hard to explain what knowledge the models have learned from human, and how the models make such decisions, which was not addressed in the literature of imitation learning. This paper addresses this research gap by proposing xGAIL, the first explainable generative adversarial imitation learning framework. The proposed xGAIL framework consists of two novel components, including Spatial Activation Maximization (SpatialAM) and Spatial Randomized Input Sampling Explanation (SpatialRISE), to extract both global and local knowledge from a well-trained GAIL model that explains how a human agent makes decisions. Especially, we take taxi drivers' passenger-seeking strategy as an example to validate the effectiveness of the proposed xGAIL framework. Our analysis on a large-scale real-world taxi trajectory data shows promisingmore »results from two aspects: i) global explainable knowledge of what nearby traffic condition impels a taxi driver to choose a particular direction to find the next passenger, and ii) local explainable knowledge of what key (sometimes hidden) factors a taxi driver considers when making a particular decision.« less
  2. Deep learning models are vulnerable to adversarial examples. Most of current adversarial attacks add pixel-wise perturbations restricted to some L^p-norm, and defense models are evaluated also on adversarial examples restricted inside L^p-norm balls. However, we wish to explore adversarial examples exist beyond L^p-norm balls and their implications for attacks and defenses. In this paper, we focus on adversarial images generated by transformations. We start with color transformation and propose two gradient-based attacks. Since L^p-norm is inappropriate for measuring image quality in the transformation space, we use the similarity between transformations and the Structural Similarity Index. Next, we explore a larger transformation space consisting of combinations of color and affine transformations. We evaluate our transformation attacks on three data sets --- CIFAR10, SVHN, and ImageNet --- and their corresponding models. Finally, we perform retraining defenses to evaluate the strength of our attacks. The results show that transformation attacks are powerful. They find high-quality adversarial images that have higher transferability and misclassification rates than C&W's L^p attacks, especially at high confidence levels. They are also significantly harder to defend against by retraining than C&W's L^p attacks. More importantly, exploring different attack spaces makes it more challenging to train a universally robust model.
  3. Raynal, Ann M. ; Ranney, Kenneth I. (Ed.)
    Most research in technologies for the Deaf community have focused on translation using either video or wearable devices. Sensor-augmented gloves have been reported to yield higher gesture recognition rates than camera-based systems; however, they cannot capture information expressed through head and body movement. Gloves are also intrusive and inhibit users in their pursuit of normal daily life, while cameras can raise concerns over privacy and are ineffective in the dark. In contrast, RF sensors are non-contact, non-invasive and do not reveal private information even if hacked. Although RF sensors are unable to measure facial expressions or hand shapes, which would be required for complete translation, this paper aims to exploit near real-time ASL recognition using RF sensors for the design of smart Deaf spaces. In this way, we hope to enable the Deaf community to benefit from advances in technologies that could generate tangible improvements in their quality of life. More specifically, this paper investigates near real-time implementation of machine learning and deep learning architectures for the purpose of sequential ASL signing recognition. We utilize a 60 GHz RF sensor which transmits a frequency modulation continuous wave (FMWC waveform). RF sensors can acquire a unique source of information that ismore »inaccessible to optical or wearable devices: namely, a visual representation of the kinematic patterns of motion via the micro-Doppler signature. Micro-Doppler refers to frequency modulations that appear about the central Doppler shift, which are caused by rotational or vibrational motions that deviate from principle translational motion. In prior work, we showed that fractal complexity computed from RF data could be used to discriminate signing from daily activities and that RF data could reveal linguistic properties, such as coarticulation. We have also shown that machine learning can be used to discriminate with 99% accuracy the signing of native Deaf ASL users from that of copysigning (or imitation signing) by hearing individuals. Therefore, imitation signing data is not effective for directly training deep models. But, adversarial learning can be used to transform imitation signing to resemble native signing, or, alternatively, physics-aware generative models can be used to synthesize ASL micro-Doppler signatures for training deep neural networks. With such approaches, we have achieved over 90% recognition accuracy of 20 ASL signs. In natural environments, however, near real-time implementations of classification algorithms are required, as well as an ability to process data streams in a continuous and sequential fashion. In this work, we focus on extensions of our prior work towards this aim, and compare the efficacy of various approaches for embedding deep neural networks (DNNs) on platforms such as a Raspberry Pi or Jetson board. We examine methods for optimizing the size and computational complexity of DNNs for embedded micro-Doppler analysis, methods for network compression, and their resulting sequential ASL recognition performance.« less
  4. Though generative adversarial networks (GANs) are prominent models to generate realistic and crisp images, they are unstable to train and suffer from the mode collapse problem. The problems of GANs come from approximating the intrinsic discontinuous distribution transform map with continuous DNNs. The recently proposed AE-OT model addresses the discontinuity problem by explicitly computing the discontinuous optimal transform map in the latent space of the autoencoder. Though have no mode collapse, the generated images by AE-OT are blurry. In this paper, we propose the AE-OT-GAN model to utilize the advantages of the both models: generate high quality images and at the same time overcome the mode collapse problems. Specifically, we firstly embed the low dimensional image manifold into the latent space by autoencoder (AE). Then the extended semi-discrete optimal transport (SDOT) map is used to generate new latent codes. Finally, our GAN model is trained to generate high quality images from the latent distribution induced by the extended SDOT map. The distribution transform map from this dataset related latent distribution to the data distribution will be continuous, and thus can be well approximated by the continuous DNNs. Additionally, the paired data between the latent codes and the real images givesmore »us further restriction about the generator and stabilizes the training process. Experiments on simple MNIST dataset and complex datasets like CIFAR10 and CelebA show the advantages of the proposed method.« less
  5. As deep neural networks (DNNs) achieve extraordi- nary performance in a wide range of tasks, testing their robust- ness under adversarial attacks becomes paramount. Adversarial attacks, also known as adversarial examples, are used to measure the robustness of DNNs and are generated by incorporating imperceptible perturbations into the input data with the intention of altering a DNN’s classification. In prior work in this area, most of the proposed optimization based methods employ gradient descent to find adversarial examples. In this paper, we present an innovative method which generates adversarial examples via convex programming. Our experiment results demonstrate that we can generate adversarial examples with lower distortion and higher transferability than the C&W attack, which is the current state-of-the-art adversarial attack method for DNNs. We achieve 100% attack success rate on both the original undefended models and the adversarially-trained models. Our distortions of the L∞ attack are respectively 31% and 18% lower than the C&W attack for the best case and average case on the CIFAR-10 data set.