Search for: All records

Adversarial training (AT) has become a popular choice for training robust networks. However, it tends to sacrifice clean accuracy heavily in favor of robustness and suffers from a large generalization error. To address these concerns, we propose Smooth Adversarial Training (SAT), guided by our analysis on the eigenspectrum of the loss Hessian. We find that curriculum learning, a scheme that emphasizes on starting “easy” and gradually ramping up on the “difficulty” of training, smooths the adversarial loss landscape for a suitably chosen difficulty metric. We present a general formulation for curriculum learning in the adversarial setting and propose two difficulty metrics based on the maximal Hessian eigenvalue (H-SAT) and the softmax probability (P-SAT). We demonstrate that SAT stabilizes network training even for a large perturbation norm and allows the network to operate at a better clean accuracy versus robustness trade-off curve compared to AT. This leads to a significant improvement in both clean accuracy and robustness compared to AT, TRADES, and other baselines. To highlight a few results, our best model improves normal and robust accuracy by 6% and 1% on CIFAR-100 compared to AT, respectively. On Imagenette, a ten-class subset of ImageNet, our model outperforms AT by 23% and 3% on normal and robust accuracy respectively. 

Public release of wrist-worn motion sensor data is growing. They enable and accelerate research in developing new algorithms to passively track daily activities, resulting in improved health and wellness utilities of smartwatches and activity trackers. But, when combined with sensitive attribute inference attack and linkage attack via re-identification of the same user in multiple datasets, undisclosed sensitive attributes can be revealed to unintended organizations with potentially adverse consequences for unsuspecting data contributing users. To guide both users and data collecting researchers, we characterize the re-identification risks inherent in motion sensor data collected from wrist-worn devices in users' natural environment. For this purpose, we use an open-set formulation, train a deep learning architecture with a new loss function, and apply our model to a new data set consisting of 10 weeks of daily sensor wearing by 353 users. We find that re-identification risk increases with an increase in the activity intensity. On average, such risk is 96% for a user when sharing a full day of sensor data. 

Edge devices rely extensively on machine learning for intelligent inferences and pattern matching. However, edge devices use a multitude of sensing modalities and are exposed to wide ranging contexts. It is difficult to develop separate machine learning models for each scenario as manual labeling is not scalable. To reduce the amount of labeled data and to speed up the training process, we propose to transfer knowledge between edge devices by using unlabeled data. Our approach, called RecycleML, uses cross modal transfer to accelerate the learning of edge devices across different sensing modalities. Using human activity recognition as a case study, over our collected CMActivity dataset, we observe that RecycleML reduces the amount of required labeled data by at least 90% and speeds up the training process by up to 50 times in comparison to training the edge device from scratch. 

Our ability to synthesize sensory data that preserves specific statistical properties of the real data has had tremendous implications on data privacy and big data analytics. The synthetic data can be used as a substitute for selective real data segments,that are sensitive to the user, thus protecting privacy and resulting in improved analytics.However, increasingly adversarial roles taken by data recipients such as mobile apps, or other cloud-based analytics services, mandate that the synthetic data, in addition to preserving statistical properties, should also be difficult to distinguish from the real data. Typically, visual inspection has been used as a test to distinguish between datasets. But more recently, sophisticated classifier models (discriminators), corresponding to a set of events, have also been employed to distinguish between synthesized and real data. The model operates on both datasets and the respective event outputs are compared for consistency. In this paper, we take a step towards generating sensory data that can pass a deep learning based discriminator model test, and make two specific contributions: first, we present a deep learning based architecture for synthesizing sensory data. This architecture comprises of a generator model, which is a stack of multiple Long-Short-Term-Memory (LSTM) networks and a Mixture Density Network. second, we use another LSTM network based discriminator model for distinguishing between the true and the synthesized data. Using a dataset of accelerometer traces, collected using smartphones of users doing their daily activities, we show that the deep learning based discriminator model can only distinguish between the real and synthesized traces with an accuracy in the neighborhood of 50%. 

Differential privacy concepts have been successfully used to protect anonymity of individuals in population-scale analysis. Sharing of mobile sensor data, especially physiological data, raise different privacy challenges, that of protecting private behaviors that can be revealed from time series of sensor data. Existing privacy mechanisms rely on noise addition and data perturbation. But the accuracy requirement on inferences drawn from physiological data, together with well-established limits within which these data values occur, render traditional privacy mechanisms inapplicable. In this work, we define a new behavioral privacy metric based on differential privacy and propose a novel data substitution mechanism to protect behavioral privacy. We evaluate the efficacy of our scheme using 660 hours of ECG, respiration, and activity data collected from 43 participants and demonstrate that it is possible to retain meaningful utility, in terms of inference accuracy (90%), while simultaneously preserving the privacy of sensitive behaviors.